src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecEwSh_VdwLJSh_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_ElecEwSh_VdwLJSh_GeomP1P1_VF_avx_256_double
  38  * Electrostatics interaction: Ewald
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecEwSh_VdwLJSh_GeomP1P1_VF_avx_256_double
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB,jnrC,jnrD;
  61     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  62     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  63     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  64     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  65     real             rcutoff_scalar;
  66     real             *shiftvec,*fshift,*x,*f;
  67     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  68     real             scratch[4*DIM];
  69     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  70     real *           vdwioffsetptr0;
  71     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  72     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  73     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  74     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  75     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
  76     real             *charge;
  77     int              nvdwtype;
  78     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  79     int              *vdwtype;
  80     real             *vdwparam;
  81     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
  82     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
  83     __m128i          ewitab;
  84     __m256d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  85     __m256d          beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
  86     real             *ewtab;
  87     __m256d          dummy_mask,cutoff_mask;
  88     __m128           tmpmask0,tmpmask1;
  89     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
  90     __m256d          one     = _mm256_set1_pd(1.0);
  91     __m256d          two     = _mm256_set1_pd(2.0);
  92     x                = xx[0];
  93     f                = ff[0];
  94
  95     nri              = nlist->nri;
  96     iinr             = nlist->iinr;
  97     jindex           = nlist->jindex;
  98     jjnr             = nlist->jjnr;
  99     shiftidx         = nlist->shift;
 100     gid              = nlist->gid;
 101     shiftvec         = fr->shift_vec[0];
 102     fshift           = fr->fshift[0];
 103     facel            = _mm256_set1_pd(fr->epsfac);
 104     charge           = mdatoms->chargeA;
 105     nvdwtype         = fr->ntype;
 106     vdwparam         = fr->nbfp;
 107     vdwtype          = mdatoms->typeA;
 108
 109     sh_ewald         = _mm256_set1_pd(fr->ic->sh_ewald);
 110     beta             = _mm256_set1_pd(fr->ic->ewaldcoeff);
 111     beta2            = _mm256_mul_pd(beta,beta);
 112     beta3            = _mm256_mul_pd(beta,beta2);
 113
 114     ewtab            = fr->ic->tabq_coul_FDV0;
 115     ewtabscale       = _mm256_set1_pd(fr->ic->tabq_scale);
 116     ewtabhalfspace   = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
 117
 118     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 119     rcutoff_scalar   = fr->rcoulomb;
 120     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 121     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 122
 123     sh_vdw_invrcut6  = _mm256_set1_pd(fr->ic->sh_invrc6);
 124     rvdw             = _mm256_set1_pd(fr->rvdw);
 125
 126     /* Avoid stupid compiler warnings */
 127     jnrA = jnrB = jnrC = jnrD = 0;
 128     j_coord_offsetA = 0;
 129     j_coord_offsetB = 0;
 130     j_coord_offsetC = 0;
 131     j_coord_offsetD = 0;
 132
 133     outeriter        = 0;
 134     inneriter        = 0;
 135
 136     for(iidx=0;iidx<4*DIM;iidx++)
 137     {
 138         scratch[iidx] = 0.0;
 139     }
 140
 141     /* Start outer loop over neighborlists */
 142     for(iidx=0; iidx<nri; iidx++)
 143     {
 144         /* Load shift vector for this list */
 145         i_shift_offset   = DIM*shiftidx[iidx];
 146
 147         /* Load limits for loop over neighbors */
 148         j_index_start    = jindex[iidx];
 149         j_index_end      = jindex[iidx+1];
 150
 151         /* Get outer coordinate index */
 152         inr              = iinr[iidx];
 153         i_coord_offset   = DIM*inr;
 154
 155         /* Load i particle coords and add shift vector */
 156         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 157
 158         fix0             = _mm256_setzero_pd();
 159         fiy0             = _mm256_setzero_pd();
 160         fiz0             = _mm256_setzero_pd();
 161
 162         /* Load parameters for i particles */
 163         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 164         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 165
 166         /* Reset potential sums */
 167         velecsum         = _mm256_setzero_pd();
 168         vvdwsum          = _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             vdwjidx0A        = 2*vdwtype[jnrA+0];
 205             vdwjidx0B        = 2*vdwtype[jnrB+0];
 206             vdwjidx0C        = 2*vdwtype[jnrC+0];
 207             vdwjidx0D        = 2*vdwtype[jnrD+0];
 208
 209             /**************************
 210              * CALCULATE INTERACTIONS *
 211              **************************/
 212
 213             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 214             {
 215
 216             r00              = _mm256_mul_pd(rsq00,rinv00);
 217
 218             /* Compute parameters for interactions between i and j atoms */
 219             qq00             = _mm256_mul_pd(iq0,jq0);
 220             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 221                                             vdwioffsetptr0+vdwjidx0B,
 222                                             vdwioffsetptr0+vdwjidx0C,
 223                                             vdwioffsetptr0+vdwjidx0D,
 224                                             &c6_00,&c12_00);
 225
 226             /* EWALD ELECTROSTATICS */
 227
 228             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 229             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 230             ewitab           = _mm256_cvttpd_epi32(ewrt);
 231             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 232             ewitab           = _mm_slli_epi32(ewitab,2);
 233             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 234             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 235             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 236             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 237             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 238             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 239             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 240             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_sub_pd(rinv00,sh_ewald),velec));
 241             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 242
 243             /* LENNARD-JONES DISPERSION/REPULSION */
 244
 245             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 246             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 247             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 248             vvdw             = _mm256_sub_pd(_mm256_mul_pd( _mm256_sub_pd(vvdw12 , _mm256_mul_pd(c12_00,_mm256_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
 249                                           _mm256_mul_pd( _mm256_sub_pd(vvdw6,_mm256_mul_pd(c6_00,sh_vdw_invrcut6)),one_sixth));
 250             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 251
 252             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 253
 254             /* Update potential sum for this i atom from the interaction with this j atom. */
 255             velec            = _mm256_and_pd(velec,cutoff_mask);
 256             velecsum         = _mm256_add_pd(velecsum,velec);
 257             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 258             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 259
 260             fscal            = _mm256_add_pd(felec,fvdw);
 261
 262             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 263
 264             /* Calculate temporary vectorial force */
 265             tx               = _mm256_mul_pd(fscal,dx00);
 266             ty               = _mm256_mul_pd(fscal,dy00);
 267             tz               = _mm256_mul_pd(fscal,dz00);
 268
 269             /* Update vectorial force */
 270             fix0             = _mm256_add_pd(fix0,tx);
 271             fiy0             = _mm256_add_pd(fiy0,ty);
 272             fiz0             = _mm256_add_pd(fiz0,tz);
 273
 274             fjptrA             = f+j_coord_offsetA;
 275             fjptrB             = f+j_coord_offsetB;
 276             fjptrC             = f+j_coord_offsetC;
 277             fjptrD             = f+j_coord_offsetD;
 278             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 279
 280             }
 281
 282             /* Inner loop uses 64 flops */
 283         }
 284
 285         if(jidx<j_index_end)
 286         {
 287
 288             /* Get j neighbor index, and coordinate index */
 289             jnrlistA         = jjnr[jidx];
 290             jnrlistB         = jjnr[jidx+1];
 291             jnrlistC         = jjnr[jidx+2];
 292             jnrlistD         = jjnr[jidx+3];
 293             /* Sign of each element will be negative for non-real atoms.
 294              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 295              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 296              */
 297             tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 298
 299             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 300             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 301             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 302
 303             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 304             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 305             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 306             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 307             j_coord_offsetA  = DIM*jnrA;
 308             j_coord_offsetB  = DIM*jnrB;
 309             j_coord_offsetC  = DIM*jnrC;
 310             j_coord_offsetD  = DIM*jnrD;
 311
 312             /* load j atom coordinates */
 313             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 314                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 315                                                  &jx0,&jy0,&jz0);
 316
 317             /* Calculate displacement vector */
 318             dx00             = _mm256_sub_pd(ix0,jx0);
 319             dy00             = _mm256_sub_pd(iy0,jy0);
 320             dz00             = _mm256_sub_pd(iz0,jz0);
 321
 322             /* Calculate squared distance and things based on it */
 323             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 324
 325             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 326
 327             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 328
 329             /* Load parameters for j particles */
 330             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 331                                                                  charge+jnrC+0,charge+jnrD+0);
 332             vdwjidx0A        = 2*vdwtype[jnrA+0];
 333             vdwjidx0B        = 2*vdwtype[jnrB+0];
 334             vdwjidx0C        = 2*vdwtype[jnrC+0];
 335             vdwjidx0D        = 2*vdwtype[jnrD+0];
 336
 337             /**************************
 338              * CALCULATE INTERACTIONS *
 339              **************************/
 340
 341             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 342             {
 343
 344             r00              = _mm256_mul_pd(rsq00,rinv00);
 345             r00              = _mm256_andnot_pd(dummy_mask,r00);
 346
 347             /* Compute parameters for interactions between i and j atoms */
 348             qq00             = _mm256_mul_pd(iq0,jq0);
 349             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 350                                             vdwioffsetptr0+vdwjidx0B,
 351                                             vdwioffsetptr0+vdwjidx0C,
 352                                             vdwioffsetptr0+vdwjidx0D,
 353                                             &c6_00,&c12_00);
 354
 355             /* EWALD ELECTROSTATICS */
 356
 357             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 358             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 359             ewitab           = _mm256_cvttpd_epi32(ewrt);
 360             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 361             ewitab           = _mm_slli_epi32(ewitab,2);
 362             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 363             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 364             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 365             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 366             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 367             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 368             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 369             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_sub_pd(rinv00,sh_ewald),velec));
 370             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 371
 372             /* LENNARD-JONES DISPERSION/REPULSION */
 373
 374             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 375             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 376             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 377             vvdw             = _mm256_sub_pd(_mm256_mul_pd( _mm256_sub_pd(vvdw12 , _mm256_mul_pd(c12_00,_mm256_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6))), one_twelfth) ,
 378                                           _mm256_mul_pd( _mm256_sub_pd(vvdw6,_mm256_mul_pd(c6_00,sh_vdw_invrcut6)),one_sixth));
 379             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 380
 381             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 382
 383             /* Update potential sum for this i atom from the interaction with this j atom. */
 384             velec            = _mm256_and_pd(velec,cutoff_mask);
 385             velec            = _mm256_andnot_pd(dummy_mask,velec);
 386             velecsum         = _mm256_add_pd(velecsum,velec);
 387             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 388             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 389             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 390
 391             fscal            = _mm256_add_pd(felec,fvdw);
 392
 393             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 394
 395             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 396
 397             /* Calculate temporary vectorial force */
 398             tx               = _mm256_mul_pd(fscal,dx00);
 399             ty               = _mm256_mul_pd(fscal,dy00);
 400             tz               = _mm256_mul_pd(fscal,dz00);
 401
 402             /* Update vectorial force */
 403             fix0             = _mm256_add_pd(fix0,tx);
 404             fiy0             = _mm256_add_pd(fiy0,ty);
 405             fiz0             = _mm256_add_pd(fiz0,tz);
 406
 407             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 408             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 409             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 410             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 411             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 412
 413             }
 414
 415             /* Inner loop uses 65 flops */
 416         }
 417
 418         /* End of innermost loop */
 419
 420         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 421                                                  f+i_coord_offset,fshift+i_shift_offset);
 422
 423         ggid                        = gid[iidx];
 424         /* Update potential energies */
 425         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 426         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 427
 428         /* Increment number of inner iterations */
 429         inneriter                  += j_index_end - j_index_start;
 430
 431         /* Outer loop uses 9 flops */
 432     }
 433
 434     /* Increment number of outer iterations */
 435     outeriter        += nri;
 436
 437     /* Update outer/inner flops */
 438
 439     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*65);
 440 }
 441 /*
 442  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJSh_GeomP1P1_F_avx_256_double
 443  * Electrostatics interaction: Ewald
 444  * VdW interaction:            LennardJones
 445  * Geometry:                   Particle-Particle
 446  * Calculate force/pot:        Force
 447  */
 448 void
 449 nb_kernel_ElecEwSh_VdwLJSh_GeomP1P1_F_avx_256_double
 450                     (t_nblist * gmx_restrict                nlist,
 451                      rvec * gmx_restrict                    xx,
 452                      rvec * gmx_restrict                    ff,
 453                      t_forcerec * gmx_restrict              fr,
 454                      t_mdatoms * gmx_restrict               mdatoms,
 455                      nb_kernel_data_t * gmx_restrict        kernel_data,
 456                      t_nrnb * gmx_restrict                  nrnb)
 457 {
 458     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 459      * just 0 for non-waters.
 460      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 461      * jnr indices corresponding to data put in the four positions in the SIMD register.
 462      */
 463     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 464     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 465     int              jnrA,jnrB,jnrC,jnrD;
 466     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 467     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 468     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 469     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 470     real             rcutoff_scalar;
 471     real             *shiftvec,*fshift,*x,*f;
 472     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 473     real             scratch[4*DIM];
 474     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 475     real *           vdwioffsetptr0;
 476     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 477     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 478     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 479     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 480     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 481     real             *charge;
 482     int              nvdwtype;
 483     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 484     int              *vdwtype;
 485     real             *vdwparam;
 486     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 487     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 488     __m128i          ewitab;
 489     __m256d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 490     __m256d          beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 491     real             *ewtab;
 492     __m256d          dummy_mask,cutoff_mask;
 493     __m128           tmpmask0,tmpmask1;
 494     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 495     __m256d          one     = _mm256_set1_pd(1.0);
 496     __m256d          two     = _mm256_set1_pd(2.0);
 497     x                = xx[0];
 498     f                = ff[0];
 499
 500     nri              = nlist->nri;
 501     iinr             = nlist->iinr;
 502     jindex           = nlist->jindex;
 503     jjnr             = nlist->jjnr;
 504     shiftidx         = nlist->shift;
 505     gid              = nlist->gid;
 506     shiftvec         = fr->shift_vec[0];
 507     fshift           = fr->fshift[0];
 508     facel            = _mm256_set1_pd(fr->epsfac);
 509     charge           = mdatoms->chargeA;
 510     nvdwtype         = fr->ntype;
 511     vdwparam         = fr->nbfp;
 512     vdwtype          = mdatoms->typeA;
 513
 514     sh_ewald         = _mm256_set1_pd(fr->ic->sh_ewald);
 515     beta             = _mm256_set1_pd(fr->ic->ewaldcoeff);
 516     beta2            = _mm256_mul_pd(beta,beta);
 517     beta3            = _mm256_mul_pd(beta,beta2);
 518
 519     ewtab            = fr->ic->tabq_coul_F;
 520     ewtabscale       = _mm256_set1_pd(fr->ic->tabq_scale);
 521     ewtabhalfspace   = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
 522
 523     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 524     rcutoff_scalar   = fr->rcoulomb;
 525     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 526     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 527
 528     sh_vdw_invrcut6  = _mm256_set1_pd(fr->ic->sh_invrc6);
 529     rvdw             = _mm256_set1_pd(fr->rvdw);
 530
 531     /* Avoid stupid compiler warnings */
 532     jnrA = jnrB = jnrC = jnrD = 0;
 533     j_coord_offsetA = 0;
 534     j_coord_offsetB = 0;
 535     j_coord_offsetC = 0;
 536     j_coord_offsetD = 0;
 537
 538     outeriter        = 0;
 539     inneriter        = 0;
 540
 541     for(iidx=0;iidx<4*DIM;iidx++)
 542     {
 543         scratch[iidx] = 0.0;
 544     }
 545
 546     /* Start outer loop over neighborlists */
 547     for(iidx=0; iidx<nri; iidx++)
 548     {
 549         /* Load shift vector for this list */
 550         i_shift_offset   = DIM*shiftidx[iidx];
 551
 552         /* Load limits for loop over neighbors */
 553         j_index_start    = jindex[iidx];
 554         j_index_end      = jindex[iidx+1];
 555
 556         /* Get outer coordinate index */
 557         inr              = iinr[iidx];
 558         i_coord_offset   = DIM*inr;
 559
 560         /* Load i particle coords and add shift vector */
 561         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 562
 563         fix0             = _mm256_setzero_pd();
 564         fiy0             = _mm256_setzero_pd();
 565         fiz0             = _mm256_setzero_pd();
 566
 567         /* Load parameters for i particles */
 568         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 569         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 570
 571         /* Start inner kernel loop */
 572         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 573         {
 574
 575             /* Get j neighbor index, and coordinate index */
 576             jnrA             = jjnr[jidx];
 577             jnrB             = jjnr[jidx+1];
 578             jnrC             = jjnr[jidx+2];
 579             jnrD             = jjnr[jidx+3];
 580             j_coord_offsetA  = DIM*jnrA;
 581             j_coord_offsetB  = DIM*jnrB;
 582             j_coord_offsetC  = DIM*jnrC;
 583             j_coord_offsetD  = DIM*jnrD;
 584
 585             /* load j atom coordinates */
 586             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 587                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 588                                                  &jx0,&jy0,&jz0);
 589
 590             /* Calculate displacement vector */
 591             dx00             = _mm256_sub_pd(ix0,jx0);
 592             dy00             = _mm256_sub_pd(iy0,jy0);
 593             dz00             = _mm256_sub_pd(iz0,jz0);
 594
 595             /* Calculate squared distance and things based on it */
 596             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 597
 598             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 599
 600             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 601
 602             /* Load parameters for j particles */
 603             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 604                                                                  charge+jnrC+0,charge+jnrD+0);
 605             vdwjidx0A        = 2*vdwtype[jnrA+0];
 606             vdwjidx0B        = 2*vdwtype[jnrB+0];
 607             vdwjidx0C        = 2*vdwtype[jnrC+0];
 608             vdwjidx0D        = 2*vdwtype[jnrD+0];
 609
 610             /**************************
 611              * CALCULATE INTERACTIONS *
 612              **************************/
 613
 614             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 615             {
 616
 617             r00              = _mm256_mul_pd(rsq00,rinv00);
 618
 619             /* Compute parameters for interactions between i and j atoms */
 620             qq00             = _mm256_mul_pd(iq0,jq0);
 621             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 622                                             vdwioffsetptr0+vdwjidx0B,
 623                                             vdwioffsetptr0+vdwjidx0C,
 624                                             vdwioffsetptr0+vdwjidx0D,
 625                                             &c6_00,&c12_00);
 626
 627             /* EWALD ELECTROSTATICS */
 628
 629             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 630             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 631             ewitab           = _mm256_cvttpd_epi32(ewrt);
 632             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 633             gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
 634                                             ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
 635                                             &ewtabF,&ewtabFn);
 636             felec            = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
 637             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 638
 639             /* LENNARD-JONES DISPERSION/REPULSION */
 640
 641             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 642             fvdw             = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
 643
 644             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 645
 646             fscal            = _mm256_add_pd(felec,fvdw);
 647
 648             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 649
 650             /* Calculate temporary vectorial force */
 651             tx               = _mm256_mul_pd(fscal,dx00);
 652             ty               = _mm256_mul_pd(fscal,dy00);
 653             tz               = _mm256_mul_pd(fscal,dz00);
 654
 655             /* Update vectorial force */
 656             fix0             = _mm256_add_pd(fix0,tx);
 657             fiy0             = _mm256_add_pd(fiy0,ty);
 658             fiz0             = _mm256_add_pd(fiz0,tz);
 659
 660             fjptrA             = f+j_coord_offsetA;
 661             fjptrB             = f+j_coord_offsetB;
 662             fjptrC             = f+j_coord_offsetC;
 663             fjptrD             = f+j_coord_offsetD;
 664             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 665
 666             }
 667
 668             /* Inner loop uses 46 flops */
 669         }
 670
 671         if(jidx<j_index_end)
 672         {
 673
 674             /* Get j neighbor index, and coordinate index */
 675             jnrlistA         = jjnr[jidx];
 676             jnrlistB         = jjnr[jidx+1];
 677             jnrlistC         = jjnr[jidx+2];
 678             jnrlistD         = jjnr[jidx+3];
 679             /* Sign of each element will be negative for non-real atoms.
 680              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 681              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 682              */
 683             tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 684
 685             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 686             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 687             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 688
 689             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 690             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 691             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 692             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 693             j_coord_offsetA  = DIM*jnrA;
 694             j_coord_offsetB  = DIM*jnrB;
 695             j_coord_offsetC  = DIM*jnrC;
 696             j_coord_offsetD  = DIM*jnrD;
 697
 698             /* load j atom coordinates */
 699             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 700                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 701                                                  &jx0,&jy0,&jz0);
 702
 703             /* Calculate displacement vector */
 704             dx00             = _mm256_sub_pd(ix0,jx0);
 705             dy00             = _mm256_sub_pd(iy0,jy0);
 706             dz00             = _mm256_sub_pd(iz0,jz0);
 707
 708             /* Calculate squared distance and things based on it */
 709             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 710
 711             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 712
 713             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 714
 715             /* Load parameters for j particles */
 716             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 717                                                                  charge+jnrC+0,charge+jnrD+0);
 718             vdwjidx0A        = 2*vdwtype[jnrA+0];
 719             vdwjidx0B        = 2*vdwtype[jnrB+0];
 720             vdwjidx0C        = 2*vdwtype[jnrC+0];
 721             vdwjidx0D        = 2*vdwtype[jnrD+0];
 722
 723             /**************************
 724              * CALCULATE INTERACTIONS *
 725              **************************/
 726
 727             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 728             {
 729
 730             r00              = _mm256_mul_pd(rsq00,rinv00);
 731             r00              = _mm256_andnot_pd(dummy_mask,r00);
 732
 733             /* Compute parameters for interactions between i and j atoms */
 734             qq00             = _mm256_mul_pd(iq0,jq0);
 735             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 736                                             vdwioffsetptr0+vdwjidx0B,
 737                                             vdwioffsetptr0+vdwjidx0C,
 738                                             vdwioffsetptr0+vdwjidx0D,
 739                                             &c6_00,&c12_00);
 740
 741             /* EWALD ELECTROSTATICS */
 742
 743             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 744             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 745             ewitab           = _mm256_cvttpd_epi32(ewrt);
 746             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 747             gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
 748                                             ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
 749                                             &ewtabF,&ewtabFn);
 750             felec            = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
 751             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 752
 753             /* LENNARD-JONES DISPERSION/REPULSION */
 754
 755             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 756             fvdw             = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
 757
 758             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 759
 760             fscal            = _mm256_add_pd(felec,fvdw);
 761
 762             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 763
 764             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 765
 766             /* Calculate temporary vectorial force */
 767             tx               = _mm256_mul_pd(fscal,dx00);
 768             ty               = _mm256_mul_pd(fscal,dy00);
 769             tz               = _mm256_mul_pd(fscal,dz00);
 770
 771             /* Update vectorial force */
 772             fix0             = _mm256_add_pd(fix0,tx);
 773             fiy0             = _mm256_add_pd(fiy0,ty);
 774             fiz0             = _mm256_add_pd(fiz0,tz);
 775
 776             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 777             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 778             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 779             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 780             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 781
 782             }
 783
 784             /* Inner loop uses 47 flops */
 785         }
 786
 787         /* End of innermost loop */
 788
 789         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 790                                                  f+i_coord_offset,fshift+i_shift_offset);
 791
 792         /* Increment number of inner iterations */
 793         inneriter                  += j_index_end - j_index_start;
 794
 795         /* Outer loop uses 7 flops */
 796     }
 797
 798     /* Increment number of outer iterations */
 799     outeriter        += nri;
 800
 801     /* Update outer/inner flops */
 802
 803     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*47);
 804 }