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) 2012,2013,2014, 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 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_256_double.h"
  48 #include "kernelutil_x86_avx_256_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJSh_GeomP1P1_VF_avx_256_double
  52  * Electrostatics interaction: Ewald
  53  * VdW interaction:            LennardJones
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecEwSh_VdwLJSh_GeomP1P1_VF_avx_256_double
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
  77     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  78     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  79     real             rcutoff_scalar;
  80     real             *shiftvec,*fshift,*x,*f;
  81     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  82     real             scratch[4*DIM];
  83     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  84     real *           vdwioffsetptr0;
  85     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  86     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  87     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  88     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  89     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
  90     real             *charge;
  91     int              nvdwtype;
  92     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  93     int              *vdwtype;
  94     real             *vdwparam;
  95     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
  96     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
  97     __m128i          ewitab;
  98     __m256d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  99     __m256d          beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 100     real             *ewtab;
 101     __m256d          dummy_mask,cutoff_mask;
 102     __m128           tmpmask0,tmpmask1;
 103     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 104     __m256d          one     = _mm256_set1_pd(1.0);
 105     __m256d          two     = _mm256_set1_pd(2.0);
 106     x                = xx[0];
 107     f                = ff[0];
 108
 109     nri              = nlist->nri;
 110     iinr             = nlist->iinr;
 111     jindex           = nlist->jindex;
 112     jjnr             = nlist->jjnr;
 113     shiftidx         = nlist->shift;
 114     gid              = nlist->gid;
 115     shiftvec         = fr->shift_vec[0];
 116     fshift           = fr->fshift[0];
 117     facel            = _mm256_set1_pd(fr->epsfac);
 118     charge           = mdatoms->chargeA;
 119     nvdwtype         = fr->ntype;
 120     vdwparam         = fr->nbfp;
 121     vdwtype          = mdatoms->typeA;
 122
 123     sh_ewald         = _mm256_set1_pd(fr->ic->sh_ewald);
 124     beta             = _mm256_set1_pd(fr->ic->ewaldcoeff_q);
 125     beta2            = _mm256_mul_pd(beta,beta);
 126     beta3            = _mm256_mul_pd(beta,beta2);
 127
 128     ewtab            = fr->ic->tabq_coul_FDV0;
 129     ewtabscale       = _mm256_set1_pd(fr->ic->tabq_scale);
 130     ewtabhalfspace   = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
 131
 132     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 133     rcutoff_scalar   = fr->rcoulomb;
 134     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 135     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 136
 137     sh_vdw_invrcut6  = _mm256_set1_pd(fr->ic->sh_invrc6);
 138     rvdw             = _mm256_set1_pd(fr->rvdw);
 139
 140     /* Avoid stupid compiler warnings */
 141     jnrA = jnrB = jnrC = jnrD = 0;
 142     j_coord_offsetA = 0;
 143     j_coord_offsetB = 0;
 144     j_coord_offsetC = 0;
 145     j_coord_offsetD = 0;
 146
 147     outeriter        = 0;
 148     inneriter        = 0;
 149
 150     for(iidx=0;iidx<4*DIM;iidx++)
 151     {
 152         scratch[iidx] = 0.0;
 153     }
 154
 155     /* Start outer loop over neighborlists */
 156     for(iidx=0; iidx<nri; iidx++)
 157     {
 158         /* Load shift vector for this list */
 159         i_shift_offset   = DIM*shiftidx[iidx];
 160
 161         /* Load limits for loop over neighbors */
 162         j_index_start    = jindex[iidx];
 163         j_index_end      = jindex[iidx+1];
 164
 165         /* Get outer coordinate index */
 166         inr              = iinr[iidx];
 167         i_coord_offset   = DIM*inr;
 168
 169         /* Load i particle coords and add shift vector */
 170         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 171
 172         fix0             = _mm256_setzero_pd();
 173         fiy0             = _mm256_setzero_pd();
 174         fiz0             = _mm256_setzero_pd();
 175
 176         /* Load parameters for i particles */
 177         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 178         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 179
 180         /* Reset potential sums */
 181         velecsum         = _mm256_setzero_pd();
 182         vvdwsum          = _mm256_setzero_pd();
 183
 184         /* Start inner kernel loop */
 185         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 186         {
 187
 188             /* Get j neighbor index, and coordinate index */
 189             jnrA             = jjnr[jidx];
 190             jnrB             = jjnr[jidx+1];
 191             jnrC             = jjnr[jidx+2];
 192             jnrD             = jjnr[jidx+3];
 193             j_coord_offsetA  = DIM*jnrA;
 194             j_coord_offsetB  = DIM*jnrB;
 195             j_coord_offsetC  = DIM*jnrC;
 196             j_coord_offsetD  = DIM*jnrD;
 197
 198             /* load j atom coordinates */
 199             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 200                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 201                                                  &jx0,&jy0,&jz0);
 202
 203             /* Calculate displacement vector */
 204             dx00             = _mm256_sub_pd(ix0,jx0);
 205             dy00             = _mm256_sub_pd(iy0,jy0);
 206             dz00             = _mm256_sub_pd(iz0,jz0);
 207
 208             /* Calculate squared distance and things based on it */
 209             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 210
 211             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 212
 213             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 214
 215             /* Load parameters for j particles */
 216             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 217                                                                  charge+jnrC+0,charge+jnrD+0);
 218             vdwjidx0A        = 2*vdwtype[jnrA+0];
 219             vdwjidx0B        = 2*vdwtype[jnrB+0];
 220             vdwjidx0C        = 2*vdwtype[jnrC+0];
 221             vdwjidx0D        = 2*vdwtype[jnrD+0];
 222
 223             /**************************
 224              * CALCULATE INTERACTIONS *
 225              **************************/
 226
 227             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 228             {
 229
 230             r00              = _mm256_mul_pd(rsq00,rinv00);
 231
 232             /* Compute parameters for interactions between i and j atoms */
 233             qq00             = _mm256_mul_pd(iq0,jq0);
 234             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 235                                             vdwioffsetptr0+vdwjidx0B,
 236                                             vdwioffsetptr0+vdwjidx0C,
 237                                             vdwioffsetptr0+vdwjidx0D,
 238                                             &c6_00,&c12_00);
 239
 240             /* EWALD ELECTROSTATICS */
 241
 242             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 243             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 244             ewitab           = _mm256_cvttpd_epi32(ewrt);
 245             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 246             ewitab           = _mm_slli_epi32(ewitab,2);
 247             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 248             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 249             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 250             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 251             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 252             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 253             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 254             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_sub_pd(rinv00,sh_ewald),velec));
 255             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 256
 257             /* LENNARD-JONES DISPERSION/REPULSION */
 258
 259             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 260             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 261             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 262             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) ,
 263                                           _mm256_mul_pd( _mm256_sub_pd(vvdw6,_mm256_mul_pd(c6_00,sh_vdw_invrcut6)),one_sixth));
 264             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 265
 266             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 267
 268             /* Update potential sum for this i atom from the interaction with this j atom. */
 269             velec            = _mm256_and_pd(velec,cutoff_mask);
 270             velecsum         = _mm256_add_pd(velecsum,velec);
 271             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 272             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 273
 274             fscal            = _mm256_add_pd(felec,fvdw);
 275
 276             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 277
 278             /* Calculate temporary vectorial force */
 279             tx               = _mm256_mul_pd(fscal,dx00);
 280             ty               = _mm256_mul_pd(fscal,dy00);
 281             tz               = _mm256_mul_pd(fscal,dz00);
 282
 283             /* Update vectorial force */
 284             fix0             = _mm256_add_pd(fix0,tx);
 285             fiy0             = _mm256_add_pd(fiy0,ty);
 286             fiz0             = _mm256_add_pd(fiz0,tz);
 287
 288             fjptrA             = f+j_coord_offsetA;
 289             fjptrB             = f+j_coord_offsetB;
 290             fjptrC             = f+j_coord_offsetC;
 291             fjptrD             = f+j_coord_offsetD;
 292             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 293
 294             }
 295
 296             /* Inner loop uses 64 flops */
 297         }
 298
 299         if(jidx<j_index_end)
 300         {
 301
 302             /* Get j neighbor index, and coordinate index */
 303             jnrlistA         = jjnr[jidx];
 304             jnrlistB         = jjnr[jidx+1];
 305             jnrlistC         = jjnr[jidx+2];
 306             jnrlistD         = jjnr[jidx+3];
 307             /* Sign of each element will be negative for non-real atoms.
 308              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 309              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 310              */
 311             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 312
 313             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 314             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 315             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 316
 317             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 318             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 319             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 320             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 321             j_coord_offsetA  = DIM*jnrA;
 322             j_coord_offsetB  = DIM*jnrB;
 323             j_coord_offsetC  = DIM*jnrC;
 324             j_coord_offsetD  = DIM*jnrD;
 325
 326             /* load j atom coordinates */
 327             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 328                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 329                                                  &jx0,&jy0,&jz0);
 330
 331             /* Calculate displacement vector */
 332             dx00             = _mm256_sub_pd(ix0,jx0);
 333             dy00             = _mm256_sub_pd(iy0,jy0);
 334             dz00             = _mm256_sub_pd(iz0,jz0);
 335
 336             /* Calculate squared distance and things based on it */
 337             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 338
 339             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 340
 341             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 342
 343             /* Load parameters for j particles */
 344             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 345                                                                  charge+jnrC+0,charge+jnrD+0);
 346             vdwjidx0A        = 2*vdwtype[jnrA+0];
 347             vdwjidx0B        = 2*vdwtype[jnrB+0];
 348             vdwjidx0C        = 2*vdwtype[jnrC+0];
 349             vdwjidx0D        = 2*vdwtype[jnrD+0];
 350
 351             /**************************
 352              * CALCULATE INTERACTIONS *
 353              **************************/
 354
 355             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 356             {
 357
 358             r00              = _mm256_mul_pd(rsq00,rinv00);
 359             r00              = _mm256_andnot_pd(dummy_mask,r00);
 360
 361             /* Compute parameters for interactions between i and j atoms */
 362             qq00             = _mm256_mul_pd(iq0,jq0);
 363             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 364                                             vdwioffsetptr0+vdwjidx0B,
 365                                             vdwioffsetptr0+vdwjidx0C,
 366                                             vdwioffsetptr0+vdwjidx0D,
 367                                             &c6_00,&c12_00);
 368
 369             /* EWALD ELECTROSTATICS */
 370
 371             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 372             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 373             ewitab           = _mm256_cvttpd_epi32(ewrt);
 374             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 375             ewitab           = _mm_slli_epi32(ewitab,2);
 376             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 377             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 378             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 379             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 380             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 381             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 382             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 383             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(_mm256_sub_pd(rinv00,sh_ewald),velec));
 384             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 385
 386             /* LENNARD-JONES DISPERSION/REPULSION */
 387
 388             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 389             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 390             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 391             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) ,
 392                                           _mm256_mul_pd( _mm256_sub_pd(vvdw6,_mm256_mul_pd(c6_00,sh_vdw_invrcut6)),one_sixth));
 393             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 394
 395             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 396
 397             /* Update potential sum for this i atom from the interaction with this j atom. */
 398             velec            = _mm256_and_pd(velec,cutoff_mask);
 399             velec            = _mm256_andnot_pd(dummy_mask,velec);
 400             velecsum         = _mm256_add_pd(velecsum,velec);
 401             vvdw             = _mm256_and_pd(vvdw,cutoff_mask);
 402             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 403             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 404
 405             fscal            = _mm256_add_pd(felec,fvdw);
 406
 407             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 408
 409             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 410
 411             /* Calculate temporary vectorial force */
 412             tx               = _mm256_mul_pd(fscal,dx00);
 413             ty               = _mm256_mul_pd(fscal,dy00);
 414             tz               = _mm256_mul_pd(fscal,dz00);
 415
 416             /* Update vectorial force */
 417             fix0             = _mm256_add_pd(fix0,tx);
 418             fiy0             = _mm256_add_pd(fiy0,ty);
 419             fiz0             = _mm256_add_pd(fiz0,tz);
 420
 421             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 422             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 423             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 424             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 425             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 426
 427             }
 428
 429             /* Inner loop uses 65 flops */
 430         }
 431
 432         /* End of innermost loop */
 433
 434         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 435                                                  f+i_coord_offset,fshift+i_shift_offset);
 436
 437         ggid                        = gid[iidx];
 438         /* Update potential energies */
 439         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 440         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 441
 442         /* Increment number of inner iterations */
 443         inneriter                  += j_index_end - j_index_start;
 444
 445         /* Outer loop uses 9 flops */
 446     }
 447
 448     /* Increment number of outer iterations */
 449     outeriter        += nri;
 450
 451     /* Update outer/inner flops */
 452
 453     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*65);
 454 }
 455 /*
 456  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJSh_GeomP1P1_F_avx_256_double
 457  * Electrostatics interaction: Ewald
 458  * VdW interaction:            LennardJones
 459  * Geometry:                   Particle-Particle
 460  * Calculate force/pot:        Force
 461  */
 462 void
 463 nb_kernel_ElecEwSh_VdwLJSh_GeomP1P1_F_avx_256_double
 464                     (t_nblist                    * gmx_restrict       nlist,
 465                      rvec                        * gmx_restrict          xx,
 466                      rvec                        * gmx_restrict          ff,
 467                      t_forcerec                  * gmx_restrict          fr,
 468                      t_mdatoms                   * gmx_restrict     mdatoms,
 469                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 470                      t_nrnb                      * gmx_restrict        nrnb)
 471 {
 472     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 473      * just 0 for non-waters.
 474      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 475      * jnr indices corresponding to data put in the four positions in the SIMD register.
 476      */
 477     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 478     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 479     int              jnrA,jnrB,jnrC,jnrD;
 480     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 481     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 482     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 483     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 484     real             rcutoff_scalar;
 485     real             *shiftvec,*fshift,*x,*f;
 486     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 487     real             scratch[4*DIM];
 488     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 489     real *           vdwioffsetptr0;
 490     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 491     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 492     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 493     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 494     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 495     real             *charge;
 496     int              nvdwtype;
 497     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 498     int              *vdwtype;
 499     real             *vdwparam;
 500     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 501     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 502     __m128i          ewitab;
 503     __m256d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 504     __m256d          beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 505     real             *ewtab;
 506     __m256d          dummy_mask,cutoff_mask;
 507     __m128           tmpmask0,tmpmask1;
 508     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 509     __m256d          one     = _mm256_set1_pd(1.0);
 510     __m256d          two     = _mm256_set1_pd(2.0);
 511     x                = xx[0];
 512     f                = ff[0];
 513
 514     nri              = nlist->nri;
 515     iinr             = nlist->iinr;
 516     jindex           = nlist->jindex;
 517     jjnr             = nlist->jjnr;
 518     shiftidx         = nlist->shift;
 519     gid              = nlist->gid;
 520     shiftvec         = fr->shift_vec[0];
 521     fshift           = fr->fshift[0];
 522     facel            = _mm256_set1_pd(fr->epsfac);
 523     charge           = mdatoms->chargeA;
 524     nvdwtype         = fr->ntype;
 525     vdwparam         = fr->nbfp;
 526     vdwtype          = mdatoms->typeA;
 527
 528     sh_ewald         = _mm256_set1_pd(fr->ic->sh_ewald);
 529     beta             = _mm256_set1_pd(fr->ic->ewaldcoeff_q);
 530     beta2            = _mm256_mul_pd(beta,beta);
 531     beta3            = _mm256_mul_pd(beta,beta2);
 532
 533     ewtab            = fr->ic->tabq_coul_F;
 534     ewtabscale       = _mm256_set1_pd(fr->ic->tabq_scale);
 535     ewtabhalfspace   = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
 536
 537     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 538     rcutoff_scalar   = fr->rcoulomb;
 539     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 540     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 541
 542     sh_vdw_invrcut6  = _mm256_set1_pd(fr->ic->sh_invrc6);
 543     rvdw             = _mm256_set1_pd(fr->rvdw);
 544
 545     /* Avoid stupid compiler warnings */
 546     jnrA = jnrB = jnrC = jnrD = 0;
 547     j_coord_offsetA = 0;
 548     j_coord_offsetB = 0;
 549     j_coord_offsetC = 0;
 550     j_coord_offsetD = 0;
 551
 552     outeriter        = 0;
 553     inneriter        = 0;
 554
 555     for(iidx=0;iidx<4*DIM;iidx++)
 556     {
 557         scratch[iidx] = 0.0;
 558     }
 559
 560     /* Start outer loop over neighborlists */
 561     for(iidx=0; iidx<nri; iidx++)
 562     {
 563         /* Load shift vector for this list */
 564         i_shift_offset   = DIM*shiftidx[iidx];
 565
 566         /* Load limits for loop over neighbors */
 567         j_index_start    = jindex[iidx];
 568         j_index_end      = jindex[iidx+1];
 569
 570         /* Get outer coordinate index */
 571         inr              = iinr[iidx];
 572         i_coord_offset   = DIM*inr;
 573
 574         /* Load i particle coords and add shift vector */
 575         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 576
 577         fix0             = _mm256_setzero_pd();
 578         fiy0             = _mm256_setzero_pd();
 579         fiz0             = _mm256_setzero_pd();
 580
 581         /* Load parameters for i particles */
 582         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 583         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 584
 585         /* Start inner kernel loop */
 586         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 587         {
 588
 589             /* Get j neighbor index, and coordinate index */
 590             jnrA             = jjnr[jidx];
 591             jnrB             = jjnr[jidx+1];
 592             jnrC             = jjnr[jidx+2];
 593             jnrD             = jjnr[jidx+3];
 594             j_coord_offsetA  = DIM*jnrA;
 595             j_coord_offsetB  = DIM*jnrB;
 596             j_coord_offsetC  = DIM*jnrC;
 597             j_coord_offsetD  = DIM*jnrD;
 598
 599             /* load j atom coordinates */
 600             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 601                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 602                                                  &jx0,&jy0,&jz0);
 603
 604             /* Calculate displacement vector */
 605             dx00             = _mm256_sub_pd(ix0,jx0);
 606             dy00             = _mm256_sub_pd(iy0,jy0);
 607             dz00             = _mm256_sub_pd(iz0,jz0);
 608
 609             /* Calculate squared distance and things based on it */
 610             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 611
 612             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 613
 614             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 615
 616             /* Load parameters for j particles */
 617             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 618                                                                  charge+jnrC+0,charge+jnrD+0);
 619             vdwjidx0A        = 2*vdwtype[jnrA+0];
 620             vdwjidx0B        = 2*vdwtype[jnrB+0];
 621             vdwjidx0C        = 2*vdwtype[jnrC+0];
 622             vdwjidx0D        = 2*vdwtype[jnrD+0];
 623
 624             /**************************
 625              * CALCULATE INTERACTIONS *
 626              **************************/
 627
 628             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 629             {
 630
 631             r00              = _mm256_mul_pd(rsq00,rinv00);
 632
 633             /* Compute parameters for interactions between i and j atoms */
 634             qq00             = _mm256_mul_pd(iq0,jq0);
 635             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 636                                             vdwioffsetptr0+vdwjidx0B,
 637                                             vdwioffsetptr0+vdwjidx0C,
 638                                             vdwioffsetptr0+vdwjidx0D,
 639                                             &c6_00,&c12_00);
 640
 641             /* EWALD ELECTROSTATICS */
 642
 643             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 644             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 645             ewitab           = _mm256_cvttpd_epi32(ewrt);
 646             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 647             gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
 648                                             ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
 649                                             &ewtabF,&ewtabFn);
 650             felec            = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
 651             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 652
 653             /* LENNARD-JONES DISPERSION/REPULSION */
 654
 655             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 656             fvdw             = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
 657
 658             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 659
 660             fscal            = _mm256_add_pd(felec,fvdw);
 661
 662             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 663
 664             /* Calculate temporary vectorial force */
 665             tx               = _mm256_mul_pd(fscal,dx00);
 666             ty               = _mm256_mul_pd(fscal,dy00);
 667             tz               = _mm256_mul_pd(fscal,dz00);
 668
 669             /* Update vectorial force */
 670             fix0             = _mm256_add_pd(fix0,tx);
 671             fiy0             = _mm256_add_pd(fiy0,ty);
 672             fiz0             = _mm256_add_pd(fiz0,tz);
 673
 674             fjptrA             = f+j_coord_offsetA;
 675             fjptrB             = f+j_coord_offsetB;
 676             fjptrC             = f+j_coord_offsetC;
 677             fjptrD             = f+j_coord_offsetD;
 678             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 679
 680             }
 681
 682             /* Inner loop uses 46 flops */
 683         }
 684
 685         if(jidx<j_index_end)
 686         {
 687
 688             /* Get j neighbor index, and coordinate index */
 689             jnrlistA         = jjnr[jidx];
 690             jnrlistB         = jjnr[jidx+1];
 691             jnrlistC         = jjnr[jidx+2];
 692             jnrlistD         = jjnr[jidx+3];
 693             /* Sign of each element will be negative for non-real atoms.
 694              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 695              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 696              */
 697             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 698
 699             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 700             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 701             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 702
 703             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 704             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 705             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 706             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 707             j_coord_offsetA  = DIM*jnrA;
 708             j_coord_offsetB  = DIM*jnrB;
 709             j_coord_offsetC  = DIM*jnrC;
 710             j_coord_offsetD  = DIM*jnrD;
 711
 712             /* load j atom coordinates */
 713             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 714                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 715                                                  &jx0,&jy0,&jz0);
 716
 717             /* Calculate displacement vector */
 718             dx00             = _mm256_sub_pd(ix0,jx0);
 719             dy00             = _mm256_sub_pd(iy0,jy0);
 720             dz00             = _mm256_sub_pd(iz0,jz0);
 721
 722             /* Calculate squared distance and things based on it */
 723             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 724
 725             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 726
 727             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 728
 729             /* Load parameters for j particles */
 730             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 731                                                                  charge+jnrC+0,charge+jnrD+0);
 732             vdwjidx0A        = 2*vdwtype[jnrA+0];
 733             vdwjidx0B        = 2*vdwtype[jnrB+0];
 734             vdwjidx0C        = 2*vdwtype[jnrC+0];
 735             vdwjidx0D        = 2*vdwtype[jnrD+0];
 736
 737             /**************************
 738              * CALCULATE INTERACTIONS *
 739              **************************/
 740
 741             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 742             {
 743
 744             r00              = _mm256_mul_pd(rsq00,rinv00);
 745             r00              = _mm256_andnot_pd(dummy_mask,r00);
 746
 747             /* Compute parameters for interactions between i and j atoms */
 748             qq00             = _mm256_mul_pd(iq0,jq0);
 749             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 750                                             vdwioffsetptr0+vdwjidx0B,
 751                                             vdwioffsetptr0+vdwjidx0C,
 752                                             vdwioffsetptr0+vdwjidx0D,
 753                                             &c6_00,&c12_00);
 754
 755             /* EWALD ELECTROSTATICS */
 756
 757             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 758             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 759             ewitab           = _mm256_cvttpd_epi32(ewrt);
 760             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 761             gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
 762                                             ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
 763                                             &ewtabF,&ewtabFn);
 764             felec            = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
 765             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 766
 767             /* LENNARD-JONES DISPERSION/REPULSION */
 768
 769             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 770             fvdw             = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
 771
 772             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 773
 774             fscal            = _mm256_add_pd(felec,fvdw);
 775
 776             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 777
 778             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 779
 780             /* Calculate temporary vectorial force */
 781             tx               = _mm256_mul_pd(fscal,dx00);
 782             ty               = _mm256_mul_pd(fscal,dy00);
 783             tz               = _mm256_mul_pd(fscal,dz00);
 784
 785             /* Update vectorial force */
 786             fix0             = _mm256_add_pd(fix0,tx);
 787             fiy0             = _mm256_add_pd(fiy0,ty);
 788             fiz0             = _mm256_add_pd(fiz0,tz);
 789
 790             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 791             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 792             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 793             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 794             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 795
 796             }
 797
 798             /* Inner loop uses 47 flops */
 799         }
 800
 801         /* End of innermost loop */
 802
 803         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 804                                                  f+i_coord_offset,fshift+i_shift_offset);
 805
 806         /* Increment number of inner iterations */
 807         inneriter                  += j_index_end - j_index_start;
 808
 809         /* Outer loop uses 7 flops */
 810     }
 811
 812     /* Increment number of outer iterations */
 813     outeriter        += nri;
 814
 815     /* Update outer/inner flops */
 816
 817     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*47);
 818 }