src/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecEwSw_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_ElecEwSw_VdwNone_GeomP1P1_VF_avx_256_double
  38  * Electrostatics interaction: Ewald
  39  * VdW interaction:            None
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecEwSw_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          ewitab;
  78     __m256d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  79     __m256d          beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
  80     real             *ewtab;
  81     __m256d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  82     real             rswitch_scalar,d_scalar;
  83     __m256d          dummy_mask,cutoff_mask;
  84     __m128           tmpmask0,tmpmask1;
  85     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
  86     __m256d          one     = _mm256_set1_pd(1.0);
  87     __m256d          two     = _mm256_set1_pd(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_pd(fr->epsfac);
 100     charge           = mdatoms->chargeA;
 101
 102     sh_ewald         = _mm256_set1_pd(fr->ic->sh_ewald);
 103     beta             = _mm256_set1_pd(fr->ic->ewaldcoeff);
 104     beta2            = _mm256_mul_pd(beta,beta);
 105     beta3            = _mm256_mul_pd(beta,beta2);
 106
 107     ewtab            = fr->ic->tabq_coul_FDV0;
 108     ewtabscale       = _mm256_set1_pd(fr->ic->tabq_scale);
 109     ewtabhalfspace   = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
 110
 111     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 112     rcutoff_scalar   = fr->rcoulomb;
 113     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 114     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 115
 116     rswitch_scalar   = fr->rcoulomb_switch;
 117     rswitch          = _mm256_set1_pd(rswitch_scalar);
 118     /* Setup switch parameters */
 119     d_scalar         = rcutoff_scalar-rswitch_scalar;
 120     d                = _mm256_set1_pd(d_scalar);
 121     swV3             = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 122     swV4             = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 123     swV5             = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 124     swF2             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 125     swF3             = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 126     swF4             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 127
 128     /* Avoid stupid compiler warnings */
 129     jnrA = jnrB = jnrC = jnrD = 0;
 130     j_coord_offsetA = 0;
 131     j_coord_offsetB = 0;
 132     j_coord_offsetC = 0;
 133     j_coord_offsetD = 0;
 134
 135     outeriter        = 0;
 136     inneriter        = 0;
 137
 138     for(iidx=0;iidx<4*DIM;iidx++)
 139     {
 140         scratch[iidx] = 0.0;
 141     }
 142
 143     /* Start outer loop over neighborlists */
 144     for(iidx=0; iidx<nri; iidx++)
 145     {
 146         /* Load shift vector for this list */
 147         i_shift_offset   = DIM*shiftidx[iidx];
 148
 149         /* Load limits for loop over neighbors */
 150         j_index_start    = jindex[iidx];
 151         j_index_end      = jindex[iidx+1];
 152
 153         /* Get outer coordinate index */
 154         inr              = iinr[iidx];
 155         i_coord_offset   = DIM*inr;
 156
 157         /* Load i particle coords and add shift vector */
 158         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 159
 160         fix0             = _mm256_setzero_pd();
 161         fiy0             = _mm256_setzero_pd();
 162         fiz0             = _mm256_setzero_pd();
 163
 164         /* Load parameters for i particles */
 165         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 166
 167         /* Reset potential sums */
 168         velecsum         = _mm256_setzero_pd();
 169
 170         /* Start inner kernel loop */
 171         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 172         {
 173
 174             /* Get j neighbor index, and coordinate index */
 175             jnrA             = jjnr[jidx];
 176             jnrB             = jjnr[jidx+1];
 177             jnrC             = jjnr[jidx+2];
 178             jnrD             = jjnr[jidx+3];
 179             j_coord_offsetA  = DIM*jnrA;
 180             j_coord_offsetB  = DIM*jnrB;
 181             j_coord_offsetC  = DIM*jnrC;
 182             j_coord_offsetD  = DIM*jnrD;
 183
 184             /* load j atom coordinates */
 185             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 186                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 187                                                  &jx0,&jy0,&jz0);
 188
 189             /* Calculate displacement vector */
 190             dx00             = _mm256_sub_pd(ix0,jx0);
 191             dy00             = _mm256_sub_pd(iy0,jy0);
 192             dz00             = _mm256_sub_pd(iz0,jz0);
 193
 194             /* Calculate squared distance and things based on it */
 195             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 196
 197             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 198
 199             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 200
 201             /* Load parameters for j particles */
 202             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 203                                                                  charge+jnrC+0,charge+jnrD+0);
 204
 205             /**************************
 206              * CALCULATE INTERACTIONS *
 207              **************************/
 208
 209             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 210             {
 211
 212             r00              = _mm256_mul_pd(rsq00,rinv00);
 213
 214             /* Compute parameters for interactions between i and j atoms */
 215             qq00             = _mm256_mul_pd(iq0,jq0);
 216
 217             /* EWALD ELECTROSTATICS */
 218
 219             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 220             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 221             ewitab           = _mm256_cvttpd_epi32(ewrt);
 222             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 223             ewitab           = _mm_slli_epi32(ewitab,2);
 224             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 225             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 226             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 227             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 228             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 229             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 230             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 231             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
 232             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 233
 234             d                = _mm256_sub_pd(r00,rswitch);
 235             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 236             d2               = _mm256_mul_pd(d,d);
 237             sw               = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
 238
 239             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 240
 241             /* Evaluate switch function */
 242             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 243             felec            = _mm256_sub_pd( _mm256_mul_pd(felec,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(velec,dsw)) );
 244             velec            = _mm256_mul_pd(velec,sw);
 245             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 246
 247             /* Update potential sum for this i atom from the interaction with this j atom. */
 248             velec            = _mm256_and_pd(velec,cutoff_mask);
 249             velecsum         = _mm256_add_pd(velecsum,velec);
 250
 251             fscal            = felec;
 252
 253             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 254
 255             /* Calculate temporary vectorial force */
 256             tx               = _mm256_mul_pd(fscal,dx00);
 257             ty               = _mm256_mul_pd(fscal,dy00);
 258             tz               = _mm256_mul_pd(fscal,dz00);
 259
 260             /* Update vectorial force */
 261             fix0             = _mm256_add_pd(fix0,tx);
 262             fiy0             = _mm256_add_pd(fiy0,ty);
 263             fiz0             = _mm256_add_pd(fiz0,tz);
 264
 265             fjptrA             = f+j_coord_offsetA;
 266             fjptrB             = f+j_coord_offsetB;
 267             fjptrC             = f+j_coord_offsetC;
 268             fjptrD             = f+j_coord_offsetD;
 269             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 270
 271             }
 272
 273             /* Inner loop uses 65 flops */
 274         }
 275
 276         if(jidx<j_index_end)
 277         {
 278
 279             /* Get j neighbor index, and coordinate index */
 280             jnrlistA         = jjnr[jidx];
 281             jnrlistB         = jjnr[jidx+1];
 282             jnrlistC         = jjnr[jidx+2];
 283             jnrlistD         = jjnr[jidx+3];
 284             /* Sign of each element will be negative for non-real atoms.
 285              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 286              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 287              */
 288             tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 289
 290             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 291             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 292             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 293
 294             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 295             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 296             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 297             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 298             j_coord_offsetA  = DIM*jnrA;
 299             j_coord_offsetB  = DIM*jnrB;
 300             j_coord_offsetC  = DIM*jnrC;
 301             j_coord_offsetD  = DIM*jnrD;
 302
 303             /* load j atom coordinates */
 304             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 305                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 306                                                  &jx0,&jy0,&jz0);
 307
 308             /* Calculate displacement vector */
 309             dx00             = _mm256_sub_pd(ix0,jx0);
 310             dy00             = _mm256_sub_pd(iy0,jy0);
 311             dz00             = _mm256_sub_pd(iz0,jz0);
 312
 313             /* Calculate squared distance and things based on it */
 314             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 315
 316             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 317
 318             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 319
 320             /* Load parameters for j particles */
 321             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 322                                                                  charge+jnrC+0,charge+jnrD+0);
 323
 324             /**************************
 325              * CALCULATE INTERACTIONS *
 326              **************************/
 327
 328             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 329             {
 330
 331             r00              = _mm256_mul_pd(rsq00,rinv00);
 332             r00              = _mm256_andnot_pd(dummy_mask,r00);
 333
 334             /* Compute parameters for interactions between i and j atoms */
 335             qq00             = _mm256_mul_pd(iq0,jq0);
 336
 337             /* EWALD ELECTROSTATICS */
 338
 339             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 340             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 341             ewitab           = _mm256_cvttpd_epi32(ewrt);
 342             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 343             ewitab           = _mm_slli_epi32(ewitab,2);
 344             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 345             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 346             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 347             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 348             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 349             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 350             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 351             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
 352             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 353
 354             d                = _mm256_sub_pd(r00,rswitch);
 355             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 356             d2               = _mm256_mul_pd(d,d);
 357             sw               = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
 358
 359             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 360
 361             /* Evaluate switch function */
 362             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 363             felec            = _mm256_sub_pd( _mm256_mul_pd(felec,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(velec,dsw)) );
 364             velec            = _mm256_mul_pd(velec,sw);
 365             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 366
 367             /* Update potential sum for this i atom from the interaction with this j atom. */
 368             velec            = _mm256_and_pd(velec,cutoff_mask);
 369             velec            = _mm256_andnot_pd(dummy_mask,velec);
 370             velecsum         = _mm256_add_pd(velecsum,velec);
 371
 372             fscal            = felec;
 373
 374             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 375
 376             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 377
 378             /* Calculate temporary vectorial force */
 379             tx               = _mm256_mul_pd(fscal,dx00);
 380             ty               = _mm256_mul_pd(fscal,dy00);
 381             tz               = _mm256_mul_pd(fscal,dz00);
 382
 383             /* Update vectorial force */
 384             fix0             = _mm256_add_pd(fix0,tx);
 385             fiy0             = _mm256_add_pd(fiy0,ty);
 386             fiz0             = _mm256_add_pd(fiz0,tz);
 387
 388             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 389             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 390             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 391             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 392             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 393
 394             }
 395
 396             /* Inner loop uses 66 flops */
 397         }
 398
 399         /* End of innermost loop */
 400
 401         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 402                                                  f+i_coord_offset,fshift+i_shift_offset);
 403
 404         ggid                        = gid[iidx];
 405         /* Update potential energies */
 406         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 407
 408         /* Increment number of inner iterations */
 409         inneriter                  += j_index_end - j_index_start;
 410
 411         /* Outer loop uses 8 flops */
 412     }
 413
 414     /* Increment number of outer iterations */
 415     outeriter        += nri;
 416
 417     /* Update outer/inner flops */
 418
 419     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*66);
 420 }
 421 /*
 422  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwNone_GeomP1P1_F_avx_256_double
 423  * Electrostatics interaction: Ewald
 424  * VdW interaction:            None
 425  * Geometry:                   Particle-Particle
 426  * Calculate force/pot:        Force
 427  */
 428 void
 429 nb_kernel_ElecEwSw_VdwNone_GeomP1P1_F_avx_256_double
 430                     (t_nblist * gmx_restrict                nlist,
 431                      rvec * gmx_restrict                    xx,
 432                      rvec * gmx_restrict                    ff,
 433                      t_forcerec * gmx_restrict              fr,
 434                      t_mdatoms * gmx_restrict               mdatoms,
 435                      nb_kernel_data_t * gmx_restrict        kernel_data,
 436                      t_nrnb * gmx_restrict                  nrnb)
 437 {
 438     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 439      * just 0 for non-waters.
 440      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 441      * jnr indices corresponding to data put in the four positions in the SIMD register.
 442      */
 443     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 444     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 445     int              jnrA,jnrB,jnrC,jnrD;
 446     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 447     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 448     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 449     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 450     real             rcutoff_scalar;
 451     real             *shiftvec,*fshift,*x,*f;
 452     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 453     real             scratch[4*DIM];
 454     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 455     real *           vdwioffsetptr0;
 456     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 457     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 458     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 459     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 460     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 461     real             *charge;
 462     __m128i          ewitab;
 463     __m256d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 464     __m256d          beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 465     real             *ewtab;
 466     __m256d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 467     real             rswitch_scalar,d_scalar;
 468     __m256d          dummy_mask,cutoff_mask;
 469     __m128           tmpmask0,tmpmask1;
 470     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 471     __m256d          one     = _mm256_set1_pd(1.0);
 472     __m256d          two     = _mm256_set1_pd(2.0);
 473     x                = xx[0];
 474     f                = ff[0];
 475
 476     nri              = nlist->nri;
 477     iinr             = nlist->iinr;
 478     jindex           = nlist->jindex;
 479     jjnr             = nlist->jjnr;
 480     shiftidx         = nlist->shift;
 481     gid              = nlist->gid;
 482     shiftvec         = fr->shift_vec[0];
 483     fshift           = fr->fshift[0];
 484     facel            = _mm256_set1_pd(fr->epsfac);
 485     charge           = mdatoms->chargeA;
 486
 487     sh_ewald         = _mm256_set1_pd(fr->ic->sh_ewald);
 488     beta             = _mm256_set1_pd(fr->ic->ewaldcoeff);
 489     beta2            = _mm256_mul_pd(beta,beta);
 490     beta3            = _mm256_mul_pd(beta,beta2);
 491
 492     ewtab            = fr->ic->tabq_coul_FDV0;
 493     ewtabscale       = _mm256_set1_pd(fr->ic->tabq_scale);
 494     ewtabhalfspace   = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
 495
 496     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 497     rcutoff_scalar   = fr->rcoulomb;
 498     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 499     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 500
 501     rswitch_scalar   = fr->rcoulomb_switch;
 502     rswitch          = _mm256_set1_pd(rswitch_scalar);
 503     /* Setup switch parameters */
 504     d_scalar         = rcutoff_scalar-rswitch_scalar;
 505     d                = _mm256_set1_pd(d_scalar);
 506     swV3             = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 507     swV4             = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 508     swV5             = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 509     swF2             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 510     swF3             = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 511     swF4             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 512
 513     /* Avoid stupid compiler warnings */
 514     jnrA = jnrB = jnrC = jnrD = 0;
 515     j_coord_offsetA = 0;
 516     j_coord_offsetB = 0;
 517     j_coord_offsetC = 0;
 518     j_coord_offsetD = 0;
 519
 520     outeriter        = 0;
 521     inneriter        = 0;
 522
 523     for(iidx=0;iidx<4*DIM;iidx++)
 524     {
 525         scratch[iidx] = 0.0;
 526     }
 527
 528     /* Start outer loop over neighborlists */
 529     for(iidx=0; iidx<nri; iidx++)
 530     {
 531         /* Load shift vector for this list */
 532         i_shift_offset   = DIM*shiftidx[iidx];
 533
 534         /* Load limits for loop over neighbors */
 535         j_index_start    = jindex[iidx];
 536         j_index_end      = jindex[iidx+1];
 537
 538         /* Get outer coordinate index */
 539         inr              = iinr[iidx];
 540         i_coord_offset   = DIM*inr;
 541
 542         /* Load i particle coords and add shift vector */
 543         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 544
 545         fix0             = _mm256_setzero_pd();
 546         fiy0             = _mm256_setzero_pd();
 547         fiz0             = _mm256_setzero_pd();
 548
 549         /* Load parameters for i particles */
 550         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 551
 552         /* Start inner kernel loop */
 553         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 554         {
 555
 556             /* Get j neighbor index, and coordinate index */
 557             jnrA             = jjnr[jidx];
 558             jnrB             = jjnr[jidx+1];
 559             jnrC             = jjnr[jidx+2];
 560             jnrD             = jjnr[jidx+3];
 561             j_coord_offsetA  = DIM*jnrA;
 562             j_coord_offsetB  = DIM*jnrB;
 563             j_coord_offsetC  = DIM*jnrC;
 564             j_coord_offsetD  = DIM*jnrD;
 565
 566             /* load j atom coordinates */
 567             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 568                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 569                                                  &jx0,&jy0,&jz0);
 570
 571             /* Calculate displacement vector */
 572             dx00             = _mm256_sub_pd(ix0,jx0);
 573             dy00             = _mm256_sub_pd(iy0,jy0);
 574             dz00             = _mm256_sub_pd(iz0,jz0);
 575
 576             /* Calculate squared distance and things based on it */
 577             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 578
 579             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 580
 581             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 582
 583             /* Load parameters for j particles */
 584             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 585                                                                  charge+jnrC+0,charge+jnrD+0);
 586
 587             /**************************
 588              * CALCULATE INTERACTIONS *
 589              **************************/
 590
 591             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 592             {
 593
 594             r00              = _mm256_mul_pd(rsq00,rinv00);
 595
 596             /* Compute parameters for interactions between i and j atoms */
 597             qq00             = _mm256_mul_pd(iq0,jq0);
 598
 599             /* EWALD ELECTROSTATICS */
 600
 601             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 602             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 603             ewitab           = _mm256_cvttpd_epi32(ewrt);
 604             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 605             ewitab           = _mm_slli_epi32(ewitab,2);
 606             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 607             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 608             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 609             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 610             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 611             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 612             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 613             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
 614             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 615
 616             d                = _mm256_sub_pd(r00,rswitch);
 617             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 618             d2               = _mm256_mul_pd(d,d);
 619             sw               = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
 620
 621             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 622
 623             /* Evaluate switch function */
 624             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 625             felec            = _mm256_sub_pd( _mm256_mul_pd(felec,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(velec,dsw)) );
 626             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 627
 628             fscal            = felec;
 629
 630             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 631
 632             /* Calculate temporary vectorial force */
 633             tx               = _mm256_mul_pd(fscal,dx00);
 634             ty               = _mm256_mul_pd(fscal,dy00);
 635             tz               = _mm256_mul_pd(fscal,dz00);
 636
 637             /* Update vectorial force */
 638             fix0             = _mm256_add_pd(fix0,tx);
 639             fiy0             = _mm256_add_pd(fiy0,ty);
 640             fiz0             = _mm256_add_pd(fiz0,tz);
 641
 642             fjptrA             = f+j_coord_offsetA;
 643             fjptrB             = f+j_coord_offsetB;
 644             fjptrC             = f+j_coord_offsetC;
 645             fjptrD             = f+j_coord_offsetD;
 646             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 647
 648             }
 649
 650             /* Inner loop uses 62 flops */
 651         }
 652
 653         if(jidx<j_index_end)
 654         {
 655
 656             /* Get j neighbor index, and coordinate index */
 657             jnrlistA         = jjnr[jidx];
 658             jnrlistB         = jjnr[jidx+1];
 659             jnrlistC         = jjnr[jidx+2];
 660             jnrlistD         = jjnr[jidx+3];
 661             /* Sign of each element will be negative for non-real atoms.
 662              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 663              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 664              */
 665             tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 666
 667             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 668             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 669             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 670
 671             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 672             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 673             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 674             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 675             j_coord_offsetA  = DIM*jnrA;
 676             j_coord_offsetB  = DIM*jnrB;
 677             j_coord_offsetC  = DIM*jnrC;
 678             j_coord_offsetD  = DIM*jnrD;
 679
 680             /* load j atom coordinates */
 681             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 682                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 683                                                  &jx0,&jy0,&jz0);
 684
 685             /* Calculate displacement vector */
 686             dx00             = _mm256_sub_pd(ix0,jx0);
 687             dy00             = _mm256_sub_pd(iy0,jy0);
 688             dz00             = _mm256_sub_pd(iz0,jz0);
 689
 690             /* Calculate squared distance and things based on it */
 691             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 692
 693             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 694
 695             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 696
 697             /* Load parameters for j particles */
 698             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 699                                                                  charge+jnrC+0,charge+jnrD+0);
 700
 701             /**************************
 702              * CALCULATE INTERACTIONS *
 703              **************************/
 704
 705             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 706             {
 707
 708             r00              = _mm256_mul_pd(rsq00,rinv00);
 709             r00              = _mm256_andnot_pd(dummy_mask,r00);
 710
 711             /* Compute parameters for interactions between i and j atoms */
 712             qq00             = _mm256_mul_pd(iq0,jq0);
 713
 714             /* EWALD ELECTROSTATICS */
 715
 716             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 717             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 718             ewitab           = _mm256_cvttpd_epi32(ewrt);
 719             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 720             ewitab           = _mm_slli_epi32(ewitab,2);
 721             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 722             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 723             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 724             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 725             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 726             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 727             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 728             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
 729             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 730
 731             d                = _mm256_sub_pd(r00,rswitch);
 732             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 733             d2               = _mm256_mul_pd(d,d);
 734             sw               = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
 735
 736             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 737
 738             /* Evaluate switch function */
 739             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 740             felec            = _mm256_sub_pd( _mm256_mul_pd(felec,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(velec,dsw)) );
 741             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 742
 743             fscal            = felec;
 744
 745             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 746
 747             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 748
 749             /* Calculate temporary vectorial force */
 750             tx               = _mm256_mul_pd(fscal,dx00);
 751             ty               = _mm256_mul_pd(fscal,dy00);
 752             tz               = _mm256_mul_pd(fscal,dz00);
 753
 754             /* Update vectorial force */
 755             fix0             = _mm256_add_pd(fix0,tx);
 756             fiy0             = _mm256_add_pd(fiy0,ty);
 757             fiz0             = _mm256_add_pd(fiz0,tz);
 758
 759             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 760             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 761             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 762             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 763             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 764
 765             }
 766
 767             /* Inner loop uses 63 flops */
 768         }
 769
 770         /* End of innermost loop */
 771
 772         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 773                                                  f+i_coord_offset,fshift+i_shift_offset);
 774
 775         /* Increment number of inner iterations */
 776         inneriter                  += j_index_end - j_index_start;
 777
 778         /* Outer loop uses 7 flops */
 779     }
 780
 781     /* Increment number of outer iterations */
 782     outeriter        += nri;
 783
 784     /* Update outer/inner flops */
 785
 786     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*63);
 787 }