src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecEwSh_VdwLJSh_GeomP1P1_avx_256_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_256_double kernel generator.
  37  */
  38 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gmx_math_x86_avx_256_double.h"
  50 #include "kernelutil_x86_avx_256_double.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJSh_GeomP1P1_VF_avx_256_double
  54  * Electrostatics interaction: Ewald
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecEwSh_VdwLJSh_GeomP1P1_VF_avx_256_double
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  79     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  80     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  81     real             rcutoff_scalar;
  82     real             *shiftvec,*fshift,*x,*f;
  83     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  84     real             scratch[4*DIM];
  85     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  86     real *           vdwioffsetptr0;
  87     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  88     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  89     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  91     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
  92     real             *charge;
  93     int              nvdwtype;
  94     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  95     int              *vdwtype;
  96     real             *vdwparam;
  97     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
  98     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
  99     __m128i          ewitab;
 100     __m256d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 101     __m256d          beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 102     real             *ewtab;
 103     __m256d          dummy_mask,cutoff_mask;
 104     __m128           tmpmask0,tmpmask1;
 105     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 106     __m256d          one     = _mm256_set1_pd(1.0);
 107     __m256d          two     = _mm256_set1_pd(2.0);
 108     x                = xx[0];
 109     f                = ff[0];
 110
 111     nri              = nlist->nri;
 112     iinr             = nlist->iinr;
 113     jindex           = nlist->jindex;
 114     jjnr             = nlist->jjnr;
 115     shiftidx         = nlist->shift;
 116     gid              = nlist->gid;
 117     shiftvec         = fr->shift_vec[0];
 118     fshift           = fr->fshift[0];
 119     facel            = _mm256_set1_pd(fr->epsfac);
 120     charge           = mdatoms->chargeA;
 121     nvdwtype         = fr->ntype;
 122     vdwparam         = fr->nbfp;
 123     vdwtype          = mdatoms->typeA;
 124
 125     sh_ewald         = _mm256_set1_pd(fr->ic->sh_ewald);
 126     beta             = _mm256_set1_pd(fr->ic->ewaldcoeff);
 127     beta2            = _mm256_mul_pd(beta,beta);
 128     beta3            = _mm256_mul_pd(beta,beta2);
 129
 130     ewtab            = fr->ic->tabq_coul_FDV0;
 131     ewtabscale       = _mm256_set1_pd(fr->ic->tabq_scale);
 132     ewtabhalfspace   = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
 133
 134     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 135     rcutoff_scalar   = fr->rcoulomb;
 136     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 137     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 138
 139     sh_vdw_invrcut6  = _mm256_set1_pd(fr->ic->sh_invrc6);
 140     rvdw             = _mm256_set1_pd(fr->rvdw);
 141
 142     /* Avoid stupid compiler warnings */
 143     jnrA = jnrB = jnrC = jnrD = 0;
 144     j_coord_offsetA = 0;
 145     j_coord_offsetB = 0;
 146     j_coord_offsetC = 0;
 147     j_coord_offsetD = 0;
 148
 149     outeriter        = 0;
 150     inneriter        = 0;
 151
 152     for(iidx=0;iidx<4*DIM;iidx++)
 153     {
 154         scratch[iidx] = 0.0;
 155     }
 156
 157     /* Start outer loop over neighborlists */
 158     for(iidx=0; iidx<nri; iidx++)
 159     {
 160         /* Load shift vector for this list */
 161         i_shift_offset   = DIM*shiftidx[iidx];
 162
 163         /* Load limits for loop over neighbors */
 164         j_index_start    = jindex[iidx];
 165         j_index_end      = jindex[iidx+1];
 166
 167         /* Get outer coordinate index */
 168         inr              = iinr[iidx];
 169         i_coord_offset   = DIM*inr;
 170
 171         /* Load i particle coords and add shift vector */
 172         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 173
 174         fix0             = _mm256_setzero_pd();
 175         fiy0             = _mm256_setzero_pd();
 176         fiz0             = _mm256_setzero_pd();
 177
 178         /* Load parameters for i particles */
 179         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 180         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 181
 182         /* Reset potential sums */
 183         velecsum         = _mm256_setzero_pd();
 184         vvdwsum          = _mm256_setzero_pd();
 185
 186         /* Start inner kernel loop */
 187         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 188         {
 189
 190             /* Get j neighbor index, and coordinate index */
 191             jnrA             = jjnr[jidx];
 192             jnrB             = jjnr[jidx+1];
 193             jnrC             = jjnr[jidx+2];
 194             jnrD             = jjnr[jidx+3];
 195             j_coord_offsetA  = DIM*jnrA;
 196             j_coord_offsetB  = DIM*jnrB;
 197             j_coord_offsetC  = DIM*jnrC;
 198             j_coord_offsetD  = DIM*jnrD;
 199
 200             /* load j atom coordinates */
 201             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 202                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 203                                                  &jx0,&jy0,&jz0);
 204
 205             /* Calculate displacement vector */
 206             dx00             = _mm256_sub_pd(ix0,jx0);
 207             dy00             = _mm256_sub_pd(iy0,jy0);
 208             dz00             = _mm256_sub_pd(iz0,jz0);
 209
 210             /* Calculate squared distance and things based on it */
 211             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 212
 213             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 214
 215             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 216
 217             /* Load parameters for j particles */
 218             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 219                                                                  charge+jnrC+0,charge+jnrD+0);
 220             vdwjidx0A        = 2*vdwtype[jnrA+0];
 221             vdwjidx0B        = 2*vdwtype[jnrB+0];
 222             vdwjidx0C        = 2*vdwtype[jnrC+0];
 223             vdwjidx0D        = 2*vdwtype[jnrD+0];
 224
 225             /**************************
 226              * CALCULATE INTERACTIONS *
 227              **************************/
 228
 229             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 230             {
 231
 232             r00              = _mm256_mul_pd(rsq00,rinv00);
 233
 234             /* Compute parameters for interactions between i and j atoms */
 235             qq00             = _mm256_mul_pd(iq0,jq0);
 236             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 237                                             vdwioffsetptr0+vdwjidx0B,
 238                                             vdwioffsetptr0+vdwjidx0C,
 239                                             vdwioffsetptr0+vdwjidx0D,
 240                                             &c6_00,&c12_00);
 241
 242             /* EWALD ELECTROSTATICS */
 243
 244             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 245             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 246             ewitab           = _mm256_cvttpd_epi32(ewrt);
 247             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 248             ewitab           = _mm_slli_epi32(ewitab,2);
 249             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 250             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 251             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 252             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 253             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 254             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 255             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 256             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_sub_pd(rinv00,sh_ewald),velec));
 257             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 258
 259             /* LENNARD-JONES DISPERSION/REPULSION */
 260
 261             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 262             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 263             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 264             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) ,
 265                                           _mm256_mul_pd( _mm256_sub_pd(vvdw6,_mm256_mul_pd(c6_00,sh_vdw_invrcut6)),one_sixth));
 266             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 267
 268             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 269
 270             /* Update potential sum for this i atom from the interaction with this j atom. */
 271             velec            = _mm256_and_pd(velec,cutoff_mask);
 272             velecsum         = _mm256_add_pd(velecsum,velec);
 273             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 274             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 275
 276             fscal            = _mm256_add_pd(felec,fvdw);
 277
 278             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 279
 280             /* Calculate temporary vectorial force */
 281             tx               = _mm256_mul_pd(fscal,dx00);
 282             ty               = _mm256_mul_pd(fscal,dy00);
 283             tz               = _mm256_mul_pd(fscal,dz00);
 284
 285             /* Update vectorial force */
 286             fix0             = _mm256_add_pd(fix0,tx);
 287             fiy0             = _mm256_add_pd(fiy0,ty);
 288             fiz0             = _mm256_add_pd(fiz0,tz);
 289
 290             fjptrA             = f+j_coord_offsetA;
 291             fjptrB             = f+j_coord_offsetB;
 292             fjptrC             = f+j_coord_offsetC;
 293             fjptrD             = f+j_coord_offsetD;
 294             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 295
 296             }
 297
 298             /* Inner loop uses 64 flops */
 299         }
 300
 301         if(jidx<j_index_end)
 302         {
 303
 304             /* Get j neighbor index, and coordinate index */
 305             jnrlistA         = jjnr[jidx];
 306             jnrlistB         = jjnr[jidx+1];
 307             jnrlistC         = jjnr[jidx+2];
 308             jnrlistD         = jjnr[jidx+3];
 309             /* Sign of each element will be negative for non-real atoms.
 310              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 311              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 312              */
 313             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 314
 315             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 316             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 317             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 318
 319             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 320             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 321             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 322             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 323             j_coord_offsetA  = DIM*jnrA;
 324             j_coord_offsetB  = DIM*jnrB;
 325             j_coord_offsetC  = DIM*jnrC;
 326             j_coord_offsetD  = DIM*jnrD;
 327
 328             /* load j atom coordinates */
 329             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 330                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 331                                                  &jx0,&jy0,&jz0);
 332
 333             /* Calculate displacement vector */
 334             dx00             = _mm256_sub_pd(ix0,jx0);
 335             dy00             = _mm256_sub_pd(iy0,jy0);
 336             dz00             = _mm256_sub_pd(iz0,jz0);
 337
 338             /* Calculate squared distance and things based on it */
 339             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 340
 341             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 342
 343             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 344
 345             /* Load parameters for j particles */
 346             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 347                                                                  charge+jnrC+0,charge+jnrD+0);
 348             vdwjidx0A        = 2*vdwtype[jnrA+0];
 349             vdwjidx0B        = 2*vdwtype[jnrB+0];
 350             vdwjidx0C        = 2*vdwtype[jnrC+0];
 351             vdwjidx0D        = 2*vdwtype[jnrD+0];
 352
 353             /**************************
 354              * CALCULATE INTERACTIONS *
 355              **************************/
 356
 357             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 358             {
 359
 360             r00              = _mm256_mul_pd(rsq00,rinv00);
 361             r00              = _mm256_andnot_pd(dummy_mask,r00);
 362
 363             /* Compute parameters for interactions between i and j atoms */
 364             qq00             = _mm256_mul_pd(iq0,jq0);
 365             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 366                                             vdwioffsetptr0+vdwjidx0B,
 367                                             vdwioffsetptr0+vdwjidx0C,
 368                                             vdwioffsetptr0+vdwjidx0D,
 369                                             &c6_00,&c12_00);
 370
 371             /* EWALD ELECTROSTATICS */
 372
 373             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 374             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 375             ewitab           = _mm256_cvttpd_epi32(ewrt);
 376             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 377             ewitab           = _mm_slli_epi32(ewitab,2);
 378             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 379             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 380             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 381             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 382             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 383             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 384             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 385             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_sub_pd(rinv00,sh_ewald),velec));
 386             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 387
 388             /* LENNARD-JONES DISPERSION/REPULSION */
 389
 390             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 391             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 392             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 393             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) ,
 394                                           _mm256_mul_pd( _mm256_sub_pd(vvdw6,_mm256_mul_pd(c6_00,sh_vdw_invrcut6)),one_sixth));
 395             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 396
 397             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 398
 399             /* Update potential sum for this i atom from the interaction with this j atom. */
 400             velec            = _mm256_and_pd(velec,cutoff_mask);
 401             velec            = _mm256_andnot_pd(dummy_mask,velec);
 402             velecsum         = _mm256_add_pd(velecsum,velec);
 403             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 404             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 405             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 406
 407             fscal            = _mm256_add_pd(felec,fvdw);
 408
 409             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 410
 411             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 412
 413             /* Calculate temporary vectorial force */
 414             tx               = _mm256_mul_pd(fscal,dx00);
 415             ty               = _mm256_mul_pd(fscal,dy00);
 416             tz               = _mm256_mul_pd(fscal,dz00);
 417
 418             /* Update vectorial force */
 419             fix0             = _mm256_add_pd(fix0,tx);
 420             fiy0             = _mm256_add_pd(fiy0,ty);
 421             fiz0             = _mm256_add_pd(fiz0,tz);
 422
 423             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 424             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 425             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 426             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 427             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 428
 429             }
 430
 431             /* Inner loop uses 65 flops */
 432         }
 433
 434         /* End of innermost loop */
 435
 436         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 437                                                  f+i_coord_offset,fshift+i_shift_offset);
 438
 439         ggid                        = gid[iidx];
 440         /* Update potential energies */
 441         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 442         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 443
 444         /* Increment number of inner iterations */
 445         inneriter                  += j_index_end - j_index_start;
 446
 447         /* Outer loop uses 9 flops */
 448     }
 449
 450     /* Increment number of outer iterations */
 451     outeriter        += nri;
 452
 453     /* Update outer/inner flops */
 454
 455     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*65);
 456 }
 457 /*
 458  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJSh_GeomP1P1_F_avx_256_double
 459  * Electrostatics interaction: Ewald
 460  * VdW interaction:            LennardJones
 461  * Geometry:                   Particle-Particle
 462  * Calculate force/pot:        Force
 463  */
 464 void
 465 nb_kernel_ElecEwSh_VdwLJSh_GeomP1P1_F_avx_256_double
 466                     (t_nblist                    * gmx_restrict       nlist,
 467                      rvec                        * gmx_restrict          xx,
 468                      rvec                        * gmx_restrict          ff,
 469                      t_forcerec                  * gmx_restrict          fr,
 470                      t_mdatoms                   * gmx_restrict     mdatoms,
 471                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 472                      t_nrnb                      * gmx_restrict        nrnb)
 473 {
 474     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 475      * just 0 for non-waters.
 476      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 477      * jnr indices corresponding to data put in the four positions in the SIMD register.
 478      */
 479     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 480     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 481     int              jnrA,jnrB,jnrC,jnrD;
 482     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 483     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 484     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 485     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 486     real             rcutoff_scalar;
 487     real             *shiftvec,*fshift,*x,*f;
 488     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 489     real             scratch[4*DIM];
 490     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 491     real *           vdwioffsetptr0;
 492     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 493     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 494     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 495     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 496     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 497     real             *charge;
 498     int              nvdwtype;
 499     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 500     int              *vdwtype;
 501     real             *vdwparam;
 502     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 503     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 504     __m128i          ewitab;
 505     __m256d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 506     __m256d          beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 507     real             *ewtab;
 508     __m256d          dummy_mask,cutoff_mask;
 509     __m128           tmpmask0,tmpmask1;
 510     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 511     __m256d          one     = _mm256_set1_pd(1.0);
 512     __m256d          two     = _mm256_set1_pd(2.0);
 513     x                = xx[0];
 514     f                = ff[0];
 515
 516     nri              = nlist->nri;
 517     iinr             = nlist->iinr;
 518     jindex           = nlist->jindex;
 519     jjnr             = nlist->jjnr;
 520     shiftidx         = nlist->shift;
 521     gid              = nlist->gid;
 522     shiftvec         = fr->shift_vec[0];
 523     fshift           = fr->fshift[0];
 524     facel            = _mm256_set1_pd(fr->epsfac);
 525     charge           = mdatoms->chargeA;
 526     nvdwtype         = fr->ntype;
 527     vdwparam         = fr->nbfp;
 528     vdwtype          = mdatoms->typeA;
 529
 530     sh_ewald         = _mm256_set1_pd(fr->ic->sh_ewald);
 531     beta             = _mm256_set1_pd(fr->ic->ewaldcoeff);
 532     beta2            = _mm256_mul_pd(beta,beta);
 533     beta3            = _mm256_mul_pd(beta,beta2);
 534
 535     ewtab            = fr->ic->tabq_coul_F;
 536     ewtabscale       = _mm256_set1_pd(fr->ic->tabq_scale);
 537     ewtabhalfspace   = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
 538
 539     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 540     rcutoff_scalar   = fr->rcoulomb;
 541     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 542     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 543
 544     sh_vdw_invrcut6  = _mm256_set1_pd(fr->ic->sh_invrc6);
 545     rvdw             = _mm256_set1_pd(fr->rvdw);
 546
 547     /* Avoid stupid compiler warnings */
 548     jnrA = jnrB = jnrC = jnrD = 0;
 549     j_coord_offsetA = 0;
 550     j_coord_offsetB = 0;
 551     j_coord_offsetC = 0;
 552     j_coord_offsetD = 0;
 553
 554     outeriter        = 0;
 555     inneriter        = 0;
 556
 557     for(iidx=0;iidx<4*DIM;iidx++)
 558     {
 559         scratch[iidx] = 0.0;
 560     }
 561
 562     /* Start outer loop over neighborlists */
 563     for(iidx=0; iidx<nri; iidx++)
 564     {
 565         /* Load shift vector for this list */
 566         i_shift_offset   = DIM*shiftidx[iidx];
 567
 568         /* Load limits for loop over neighbors */
 569         j_index_start    = jindex[iidx];
 570         j_index_end      = jindex[iidx+1];
 571
 572         /* Get outer coordinate index */
 573         inr              = iinr[iidx];
 574         i_coord_offset   = DIM*inr;
 575
 576         /* Load i particle coords and add shift vector */
 577         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 578
 579         fix0             = _mm256_setzero_pd();
 580         fiy0             = _mm256_setzero_pd();
 581         fiz0             = _mm256_setzero_pd();
 582
 583         /* Load parameters for i particles */
 584         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 585         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 586
 587         /* Start inner kernel loop */
 588         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 589         {
 590
 591             /* Get j neighbor index, and coordinate index */
 592             jnrA             = jjnr[jidx];
 593             jnrB             = jjnr[jidx+1];
 594             jnrC             = jjnr[jidx+2];
 595             jnrD             = jjnr[jidx+3];
 596             j_coord_offsetA  = DIM*jnrA;
 597             j_coord_offsetB  = DIM*jnrB;
 598             j_coord_offsetC  = DIM*jnrC;
 599             j_coord_offsetD  = DIM*jnrD;
 600
 601             /* load j atom coordinates */
 602             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 603                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 604                                                  &jx0,&jy0,&jz0);
 605
 606             /* Calculate displacement vector */
 607             dx00             = _mm256_sub_pd(ix0,jx0);
 608             dy00             = _mm256_sub_pd(iy0,jy0);
 609             dz00             = _mm256_sub_pd(iz0,jz0);
 610
 611             /* Calculate squared distance and things based on it */
 612             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 613
 614             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 615
 616             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 617
 618             /* Load parameters for j particles */
 619             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 620                                                                  charge+jnrC+0,charge+jnrD+0);
 621             vdwjidx0A        = 2*vdwtype[jnrA+0];
 622             vdwjidx0B        = 2*vdwtype[jnrB+0];
 623             vdwjidx0C        = 2*vdwtype[jnrC+0];
 624             vdwjidx0D        = 2*vdwtype[jnrD+0];
 625
 626             /**************************
 627              * CALCULATE INTERACTIONS *
 628              **************************/
 629
 630             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 631             {
 632
 633             r00              = _mm256_mul_pd(rsq00,rinv00);
 634
 635             /* Compute parameters for interactions between i and j atoms */
 636             qq00             = _mm256_mul_pd(iq0,jq0);
 637             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 638                                             vdwioffsetptr0+vdwjidx0B,
 639                                             vdwioffsetptr0+vdwjidx0C,
 640                                             vdwioffsetptr0+vdwjidx0D,
 641                                             &c6_00,&c12_00);
 642
 643             /* EWALD ELECTROSTATICS */
 644
 645             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 646             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 647             ewitab           = _mm256_cvttpd_epi32(ewrt);
 648             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 649             gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
 650                                             ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
 651                                             &ewtabF,&ewtabFn);
 652             felec            = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
 653             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 654
 655             /* LENNARD-JONES DISPERSION/REPULSION */
 656
 657             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 658             fvdw             = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
 659
 660             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 661
 662             fscal            = _mm256_add_pd(felec,fvdw);
 663
 664             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 665
 666             /* Calculate temporary vectorial force */
 667             tx               = _mm256_mul_pd(fscal,dx00);
 668             ty               = _mm256_mul_pd(fscal,dy00);
 669             tz               = _mm256_mul_pd(fscal,dz00);
 670
 671             /* Update vectorial force */
 672             fix0             = _mm256_add_pd(fix0,tx);
 673             fiy0             = _mm256_add_pd(fiy0,ty);
 674             fiz0             = _mm256_add_pd(fiz0,tz);
 675
 676             fjptrA             = f+j_coord_offsetA;
 677             fjptrB             = f+j_coord_offsetB;
 678             fjptrC             = f+j_coord_offsetC;
 679             fjptrD             = f+j_coord_offsetD;
 680             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 681
 682             }
 683
 684             /* Inner loop uses 46 flops */
 685         }
 686
 687         if(jidx<j_index_end)
 688         {
 689
 690             /* Get j neighbor index, and coordinate index */
 691             jnrlistA         = jjnr[jidx];
 692             jnrlistB         = jjnr[jidx+1];
 693             jnrlistC         = jjnr[jidx+2];
 694             jnrlistD         = jjnr[jidx+3];
 695             /* Sign of each element will be negative for non-real atoms.
 696              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 697              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 698              */
 699             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 700
 701             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 702             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 703             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 704
 705             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 706             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 707             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 708             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 709             j_coord_offsetA  = DIM*jnrA;
 710             j_coord_offsetB  = DIM*jnrB;
 711             j_coord_offsetC  = DIM*jnrC;
 712             j_coord_offsetD  = DIM*jnrD;
 713
 714             /* load j atom coordinates */
 715             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 716                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 717                                                  &jx0,&jy0,&jz0);
 718
 719             /* Calculate displacement vector */
 720             dx00             = _mm256_sub_pd(ix0,jx0);
 721             dy00             = _mm256_sub_pd(iy0,jy0);
 722             dz00             = _mm256_sub_pd(iz0,jz0);
 723
 724             /* Calculate squared distance and things based on it */
 725             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 726
 727             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 728
 729             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 730
 731             /* Load parameters for j particles */
 732             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 733                                                                  charge+jnrC+0,charge+jnrD+0);
 734             vdwjidx0A        = 2*vdwtype[jnrA+0];
 735             vdwjidx0B        = 2*vdwtype[jnrB+0];
 736             vdwjidx0C        = 2*vdwtype[jnrC+0];
 737             vdwjidx0D        = 2*vdwtype[jnrD+0];
 738
 739             /**************************
 740              * CALCULATE INTERACTIONS *
 741              **************************/
 742
 743             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 744             {
 745
 746             r00              = _mm256_mul_pd(rsq00,rinv00);
 747             r00              = _mm256_andnot_pd(dummy_mask,r00);
 748
 749             /* Compute parameters for interactions between i and j atoms */
 750             qq00             = _mm256_mul_pd(iq0,jq0);
 751             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 752                                             vdwioffsetptr0+vdwjidx0B,
 753                                             vdwioffsetptr0+vdwjidx0C,
 754                                             vdwioffsetptr0+vdwjidx0D,
 755                                             &c6_00,&c12_00);
 756
 757             /* EWALD ELECTROSTATICS */
 758
 759             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 760             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 761             ewitab           = _mm256_cvttpd_epi32(ewrt);
 762             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 763             gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
 764                                             ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
 765                                             &ewtabF,&ewtabFn);
 766             felec            = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
 767             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 768
 769             /* LENNARD-JONES DISPERSION/REPULSION */
 770
 771             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 772             fvdw             = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
 773
 774             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 775
 776             fscal            = _mm256_add_pd(felec,fvdw);
 777
 778             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 779
 780             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 781
 782             /* Calculate temporary vectorial force */
 783             tx               = _mm256_mul_pd(fscal,dx00);
 784             ty               = _mm256_mul_pd(fscal,dy00);
 785             tz               = _mm256_mul_pd(fscal,dz00);
 786
 787             /* Update vectorial force */
 788             fix0             = _mm256_add_pd(fix0,tx);
 789             fiy0             = _mm256_add_pd(fiy0,ty);
 790             fiz0             = _mm256_add_pd(fiz0,tz);
 791
 792             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 793             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 794             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 795             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 796             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 797
 798             }
 799
 800             /* Inner loop uses 47 flops */
 801         }
 802
 803         /* End of innermost loop */
 804
 805         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 806                                                  f+i_coord_offset,fshift+i_shift_offset);
 807
 808         /* Increment number of inner iterations */
 809         inneriter                  += j_index_end - j_index_start;
 810
 811         /* Outer loop uses 7 flops */
 812     }
 813
 814     /* Increment number of outer iterations */
 815     outeriter        += nri;
 816
 817     /* Update outer/inner flops */
 818
 819     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*47);
 820 }