src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecEw_VdwLJ_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_ElecEw_VdwLJ_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_ElecEw_VdwLJ_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     /* Avoid stupid compiler warnings */
 133     jnrA = jnrB = jnrC = jnrD = 0;
 134     j_coord_offsetA = 0;
 135     j_coord_offsetB = 0;
 136     j_coord_offsetC = 0;
 137     j_coord_offsetD = 0;
 138
 139     outeriter        = 0;
 140     inneriter        = 0;
 141
 142     for(iidx=0;iidx<4*DIM;iidx++)
 143     {
 144         scratch[iidx] = 0.0;
 145     }
 146
 147     /* Start outer loop over neighborlists */
 148     for(iidx=0; iidx<nri; iidx++)
 149     {
 150         /* Load shift vector for this list */
 151         i_shift_offset   = DIM*shiftidx[iidx];
 152
 153         /* Load limits for loop over neighbors */
 154         j_index_start    = jindex[iidx];
 155         j_index_end      = jindex[iidx+1];
 156
 157         /* Get outer coordinate index */
 158         inr              = iinr[iidx];
 159         i_coord_offset   = DIM*inr;
 160
 161         /* Load i particle coords and add shift vector */
 162         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 163
 164         fix0             = _mm256_setzero_pd();
 165         fiy0             = _mm256_setzero_pd();
 166         fiz0             = _mm256_setzero_pd();
 167
 168         /* Load parameters for i particles */
 169         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 170         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 171
 172         /* Reset potential sums */
 173         velecsum         = _mm256_setzero_pd();
 174         vvdwsum          = _mm256_setzero_pd();
 175
 176         /* Start inner kernel loop */
 177         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 178         {
 179
 180             /* Get j neighbor index, and coordinate index */
 181             jnrA             = jjnr[jidx];
 182             jnrB             = jjnr[jidx+1];
 183             jnrC             = jjnr[jidx+2];
 184             jnrD             = jjnr[jidx+3];
 185             j_coord_offsetA  = DIM*jnrA;
 186             j_coord_offsetB  = DIM*jnrB;
 187             j_coord_offsetC  = DIM*jnrC;
 188             j_coord_offsetD  = DIM*jnrD;
 189
 190             /* load j atom coordinates */
 191             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 192                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 193                                                  &jx0,&jy0,&jz0);
 194
 195             /* Calculate displacement vector */
 196             dx00             = _mm256_sub_pd(ix0,jx0);
 197             dy00             = _mm256_sub_pd(iy0,jy0);
 198             dz00             = _mm256_sub_pd(iz0,jz0);
 199
 200             /* Calculate squared distance and things based on it */
 201             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 202
 203             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 204
 205             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 206
 207             /* Load parameters for j particles */
 208             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 209                                                                  charge+jnrC+0,charge+jnrD+0);
 210             vdwjidx0A        = 2*vdwtype[jnrA+0];
 211             vdwjidx0B        = 2*vdwtype[jnrB+0];
 212             vdwjidx0C        = 2*vdwtype[jnrC+0];
 213             vdwjidx0D        = 2*vdwtype[jnrD+0];
 214
 215             /**************************
 216              * CALCULATE INTERACTIONS *
 217              **************************/
 218
 219             r00              = _mm256_mul_pd(rsq00,rinv00);
 220
 221             /* Compute parameters for interactions between i and j atoms */
 222             qq00             = _mm256_mul_pd(iq0,jq0);
 223             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 224                                             vdwioffsetptr0+vdwjidx0B,
 225                                             vdwioffsetptr0+vdwjidx0C,
 226                                             vdwioffsetptr0+vdwjidx0D,
 227                                             &c6_00,&c12_00);
 228
 229             /* EWALD ELECTROSTATICS */
 230
 231             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 232             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 233             ewitab           = _mm256_cvttpd_epi32(ewrt);
 234             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 235             ewitab           = _mm_slli_epi32(ewitab,2);
 236             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 237             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 238             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 239             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 240             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 241             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 242             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 243             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
 244             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 245
 246             /* LENNARD-JONES DISPERSION/REPULSION */
 247
 248             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 249             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 250             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 251             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 252             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 253
 254             /* Update potential sum for this i atom from the interaction with this j atom. */
 255             velecsum         = _mm256_add_pd(velecsum,velec);
 256             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 257
 258             fscal            = _mm256_add_pd(felec,fvdw);
 259
 260             /* Calculate temporary vectorial force */
 261             tx               = _mm256_mul_pd(fscal,dx00);
 262             ty               = _mm256_mul_pd(fscal,dy00);
 263             tz               = _mm256_mul_pd(fscal,dz00);
 264
 265             /* Update vectorial force */
 266             fix0             = _mm256_add_pd(fix0,tx);
 267             fiy0             = _mm256_add_pd(fiy0,ty);
 268             fiz0             = _mm256_add_pd(fiz0,tz);
 269
 270             fjptrA             = f+j_coord_offsetA;
 271             fjptrB             = f+j_coord_offsetB;
 272             fjptrC             = f+j_coord_offsetC;
 273             fjptrD             = f+j_coord_offsetD;
 274             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 275
 276             /* Inner loop uses 53 flops */
 277         }
 278
 279         if(jidx<j_index_end)
 280         {
 281
 282             /* Get j neighbor index, and coordinate index */
 283             jnrlistA         = jjnr[jidx];
 284             jnrlistB         = jjnr[jidx+1];
 285             jnrlistC         = jjnr[jidx+2];
 286             jnrlistD         = jjnr[jidx+3];
 287             /* Sign of each element will be negative for non-real atoms.
 288              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 289              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 290              */
 291             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 292
 293             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 294             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 295             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 296
 297             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 298             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 299             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 300             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 301             j_coord_offsetA  = DIM*jnrA;
 302             j_coord_offsetB  = DIM*jnrB;
 303             j_coord_offsetC  = DIM*jnrC;
 304             j_coord_offsetD  = DIM*jnrD;
 305
 306             /* load j atom coordinates */
 307             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 308                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 309                                                  &jx0,&jy0,&jz0);
 310
 311             /* Calculate displacement vector */
 312             dx00             = _mm256_sub_pd(ix0,jx0);
 313             dy00             = _mm256_sub_pd(iy0,jy0);
 314             dz00             = _mm256_sub_pd(iz0,jz0);
 315
 316             /* Calculate squared distance and things based on it */
 317             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 318
 319             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 320
 321             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 322
 323             /* Load parameters for j particles */
 324             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 325                                                                  charge+jnrC+0,charge+jnrD+0);
 326             vdwjidx0A        = 2*vdwtype[jnrA+0];
 327             vdwjidx0B        = 2*vdwtype[jnrB+0];
 328             vdwjidx0C        = 2*vdwtype[jnrC+0];
 329             vdwjidx0D        = 2*vdwtype[jnrD+0];
 330
 331             /**************************
 332              * CALCULATE INTERACTIONS *
 333              **************************/
 334
 335             r00              = _mm256_mul_pd(rsq00,rinv00);
 336             r00              = _mm256_andnot_pd(dummy_mask,r00);
 337
 338             /* Compute parameters for interactions between i and j atoms */
 339             qq00             = _mm256_mul_pd(iq0,jq0);
 340             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 341                                             vdwioffsetptr0+vdwjidx0B,
 342                                             vdwioffsetptr0+vdwjidx0C,
 343                                             vdwioffsetptr0+vdwjidx0D,
 344                                             &c6_00,&c12_00);
 345
 346             /* EWALD ELECTROSTATICS */
 347
 348             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 349             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 350             ewitab           = _mm256_cvttpd_epi32(ewrt);
 351             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 352             ewitab           = _mm_slli_epi32(ewitab,2);
 353             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 354             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 355             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 356             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 357             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 358             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 359             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 360             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
 361             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 362
 363             /* LENNARD-JONES DISPERSION/REPULSION */
 364
 365             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 366             vvdw6            = _mm256_mul_pd(c6_00,rinvsix);
 367             vvdw12           = _mm256_mul_pd(c12_00,_mm256_mul_pd(rinvsix,rinvsix));
 368             vvdw             = _mm256_sub_pd( _mm256_mul_pd(vvdw12,one_twelfth) , _mm256_mul_pd(vvdw6,one_sixth) );
 369             fvdw             = _mm256_mul_pd(_mm256_sub_pd(vvdw12,vvdw6),rinvsq00);
 370
 371             /* Update potential sum for this i atom from the interaction with this j atom. */
 372             velec            = _mm256_andnot_pd(dummy_mask,velec);
 373             velecsum         = _mm256_add_pd(velecsum,velec);
 374             vvdw             = _mm256_andnot_pd(dummy_mask,vvdw);
 375             vvdwsum          = _mm256_add_pd(vvdwsum,vvdw);
 376
 377             fscal            = _mm256_add_pd(felec,fvdw);
 378
 379             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 380
 381             /* Calculate temporary vectorial force */
 382             tx               = _mm256_mul_pd(fscal,dx00);
 383             ty               = _mm256_mul_pd(fscal,dy00);
 384             tz               = _mm256_mul_pd(fscal,dz00);
 385
 386             /* Update vectorial force */
 387             fix0             = _mm256_add_pd(fix0,tx);
 388             fiy0             = _mm256_add_pd(fiy0,ty);
 389             fiz0             = _mm256_add_pd(fiz0,tz);
 390
 391             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 392             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 393             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 394             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 395             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 396
 397             /* Inner loop uses 54 flops */
 398         }
 399
 400         /* End of innermost loop */
 401
 402         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 403                                                  f+i_coord_offset,fshift+i_shift_offset);
 404
 405         ggid                        = gid[iidx];
 406         /* Update potential energies */
 407         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 408         gmx_mm256_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 409
 410         /* Increment number of inner iterations */
 411         inneriter                  += j_index_end - j_index_start;
 412
 413         /* Outer loop uses 9 flops */
 414     }
 415
 416     /* Increment number of outer iterations */
 417     outeriter        += nri;
 418
 419     /* Update outer/inner flops */
 420
 421     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*54);
 422 }
 423 /*
 424  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJ_GeomP1P1_F_avx_256_double
 425  * Electrostatics interaction: Ewald
 426  * VdW interaction:            LennardJones
 427  * Geometry:                   Particle-Particle
 428  * Calculate force/pot:        Force
 429  */
 430 void
 431 nb_kernel_ElecEw_VdwLJ_GeomP1P1_F_avx_256_double
 432                     (t_nblist                    * gmx_restrict       nlist,
 433                      rvec                        * gmx_restrict          xx,
 434                      rvec                        * gmx_restrict          ff,
 435                      t_forcerec                  * gmx_restrict          fr,
 436                      t_mdatoms                   * gmx_restrict     mdatoms,
 437                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 438                      t_nrnb                      * gmx_restrict        nrnb)
 439 {
 440     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 441      * just 0 for non-waters.
 442      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 443      * jnr indices corresponding to data put in the four positions in the SIMD register.
 444      */
 445     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 446     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 447     int              jnrA,jnrB,jnrC,jnrD;
 448     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 449     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 450     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 451     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 452     real             rcutoff_scalar;
 453     real             *shiftvec,*fshift,*x,*f;
 454     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 455     real             scratch[4*DIM];
 456     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 457     real *           vdwioffsetptr0;
 458     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 459     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 460     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 461     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 462     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 463     real             *charge;
 464     int              nvdwtype;
 465     __m256d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 466     int              *vdwtype;
 467     real             *vdwparam;
 468     __m256d          one_sixth   = _mm256_set1_pd(1.0/6.0);
 469     __m256d          one_twelfth = _mm256_set1_pd(1.0/12.0);
 470     __m128i          ewitab;
 471     __m256d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 472     __m256d          beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 473     real             *ewtab;
 474     __m256d          dummy_mask,cutoff_mask;
 475     __m128           tmpmask0,tmpmask1;
 476     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 477     __m256d          one     = _mm256_set1_pd(1.0);
 478     __m256d          two     = _mm256_set1_pd(2.0);
 479     x                = xx[0];
 480     f                = ff[0];
 481
 482     nri              = nlist->nri;
 483     iinr             = nlist->iinr;
 484     jindex           = nlist->jindex;
 485     jjnr             = nlist->jjnr;
 486     shiftidx         = nlist->shift;
 487     gid              = nlist->gid;
 488     shiftvec         = fr->shift_vec[0];
 489     fshift           = fr->fshift[0];
 490     facel            = _mm256_set1_pd(fr->epsfac);
 491     charge           = mdatoms->chargeA;
 492     nvdwtype         = fr->ntype;
 493     vdwparam         = fr->nbfp;
 494     vdwtype          = mdatoms->typeA;
 495
 496     sh_ewald         = _mm256_set1_pd(fr->ic->sh_ewald);
 497     beta             = _mm256_set1_pd(fr->ic->ewaldcoeff_q);
 498     beta2            = _mm256_mul_pd(beta,beta);
 499     beta3            = _mm256_mul_pd(beta,beta2);
 500
 501     ewtab            = fr->ic->tabq_coul_F;
 502     ewtabscale       = _mm256_set1_pd(fr->ic->tabq_scale);
 503     ewtabhalfspace   = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
 504
 505     /* Avoid stupid compiler warnings */
 506     jnrA = jnrB = jnrC = jnrD = 0;
 507     j_coord_offsetA = 0;
 508     j_coord_offsetB = 0;
 509     j_coord_offsetC = 0;
 510     j_coord_offsetD = 0;
 511
 512     outeriter        = 0;
 513     inneriter        = 0;
 514
 515     for(iidx=0;iidx<4*DIM;iidx++)
 516     {
 517         scratch[iidx] = 0.0;
 518     }
 519
 520     /* Start outer loop over neighborlists */
 521     for(iidx=0; iidx<nri; iidx++)
 522     {
 523         /* Load shift vector for this list */
 524         i_shift_offset   = DIM*shiftidx[iidx];
 525
 526         /* Load limits for loop over neighbors */
 527         j_index_start    = jindex[iidx];
 528         j_index_end      = jindex[iidx+1];
 529
 530         /* Get outer coordinate index */
 531         inr              = iinr[iidx];
 532         i_coord_offset   = DIM*inr;
 533
 534         /* Load i particle coords and add shift vector */
 535         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 536
 537         fix0             = _mm256_setzero_pd();
 538         fiy0             = _mm256_setzero_pd();
 539         fiz0             = _mm256_setzero_pd();
 540
 541         /* Load parameters for i particles */
 542         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 543         vdwioffsetptr0   = vdwparam+2*nvdwtype*vdwtype[inr+0];
 544
 545         /* Start inner kernel loop */
 546         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 547         {
 548
 549             /* Get j neighbor index, and coordinate index */
 550             jnrA             = jjnr[jidx];
 551             jnrB             = jjnr[jidx+1];
 552             jnrC             = jjnr[jidx+2];
 553             jnrD             = jjnr[jidx+3];
 554             j_coord_offsetA  = DIM*jnrA;
 555             j_coord_offsetB  = DIM*jnrB;
 556             j_coord_offsetC  = DIM*jnrC;
 557             j_coord_offsetD  = DIM*jnrD;
 558
 559             /* load j atom coordinates */
 560             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 561                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 562                                                  &jx0,&jy0,&jz0);
 563
 564             /* Calculate displacement vector */
 565             dx00             = _mm256_sub_pd(ix0,jx0);
 566             dy00             = _mm256_sub_pd(iy0,jy0);
 567             dz00             = _mm256_sub_pd(iz0,jz0);
 568
 569             /* Calculate squared distance and things based on it */
 570             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 571
 572             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 573
 574             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 575
 576             /* Load parameters for j particles */
 577             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 578                                                                  charge+jnrC+0,charge+jnrD+0);
 579             vdwjidx0A        = 2*vdwtype[jnrA+0];
 580             vdwjidx0B        = 2*vdwtype[jnrB+0];
 581             vdwjidx0C        = 2*vdwtype[jnrC+0];
 582             vdwjidx0D        = 2*vdwtype[jnrD+0];
 583
 584             /**************************
 585              * CALCULATE INTERACTIONS *
 586              **************************/
 587
 588             r00              = _mm256_mul_pd(rsq00,rinv00);
 589
 590             /* Compute parameters for interactions between i and j atoms */
 591             qq00             = _mm256_mul_pd(iq0,jq0);
 592             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 593                                             vdwioffsetptr0+vdwjidx0B,
 594                                             vdwioffsetptr0+vdwjidx0C,
 595                                             vdwioffsetptr0+vdwjidx0D,
 596                                             &c6_00,&c12_00);
 597
 598             /* EWALD ELECTROSTATICS */
 599
 600             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 601             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 602             ewitab           = _mm256_cvttpd_epi32(ewrt);
 603             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 604             gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
 605                                             ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
 606                                             &ewtabF,&ewtabFn);
 607             felec            = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
 608             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 609
 610             /* LENNARD-JONES DISPERSION/REPULSION */
 611
 612             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 613             fvdw             = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
 614
 615             fscal            = _mm256_add_pd(felec,fvdw);
 616
 617             /* Calculate temporary vectorial force */
 618             tx               = _mm256_mul_pd(fscal,dx00);
 619             ty               = _mm256_mul_pd(fscal,dy00);
 620             tz               = _mm256_mul_pd(fscal,dz00);
 621
 622             /* Update vectorial force */
 623             fix0             = _mm256_add_pd(fix0,tx);
 624             fiy0             = _mm256_add_pd(fiy0,ty);
 625             fiz0             = _mm256_add_pd(fiz0,tz);
 626
 627             fjptrA             = f+j_coord_offsetA;
 628             fjptrB             = f+j_coord_offsetB;
 629             fjptrC             = f+j_coord_offsetC;
 630             fjptrD             = f+j_coord_offsetD;
 631             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 632
 633             /* Inner loop uses 43 flops */
 634         }
 635
 636         if(jidx<j_index_end)
 637         {
 638
 639             /* Get j neighbor index, and coordinate index */
 640             jnrlistA         = jjnr[jidx];
 641             jnrlistB         = jjnr[jidx+1];
 642             jnrlistC         = jjnr[jidx+2];
 643             jnrlistD         = jjnr[jidx+3];
 644             /* Sign of each element will be negative for non-real atoms.
 645              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 646              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 647              */
 648             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 649
 650             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 651             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 652             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 653
 654             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 655             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 656             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 657             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 658             j_coord_offsetA  = DIM*jnrA;
 659             j_coord_offsetB  = DIM*jnrB;
 660             j_coord_offsetC  = DIM*jnrC;
 661             j_coord_offsetD  = DIM*jnrD;
 662
 663             /* load j atom coordinates */
 664             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 665                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 666                                                  &jx0,&jy0,&jz0);
 667
 668             /* Calculate displacement vector */
 669             dx00             = _mm256_sub_pd(ix0,jx0);
 670             dy00             = _mm256_sub_pd(iy0,jy0);
 671             dz00             = _mm256_sub_pd(iz0,jz0);
 672
 673             /* Calculate squared distance and things based on it */
 674             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 675
 676             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 677
 678             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 679
 680             /* Load parameters for j particles */
 681             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 682                                                                  charge+jnrC+0,charge+jnrD+0);
 683             vdwjidx0A        = 2*vdwtype[jnrA+0];
 684             vdwjidx0B        = 2*vdwtype[jnrB+0];
 685             vdwjidx0C        = 2*vdwtype[jnrC+0];
 686             vdwjidx0D        = 2*vdwtype[jnrD+0];
 687
 688             /**************************
 689              * CALCULATE INTERACTIONS *
 690              **************************/
 691
 692             r00              = _mm256_mul_pd(rsq00,rinv00);
 693             r00              = _mm256_andnot_pd(dummy_mask,r00);
 694
 695             /* Compute parameters for interactions between i and j atoms */
 696             qq00             = _mm256_mul_pd(iq0,jq0);
 697             gmx_mm256_load_4pair_swizzle_pd(vdwioffsetptr0+vdwjidx0A,
 698                                             vdwioffsetptr0+vdwjidx0B,
 699                                             vdwioffsetptr0+vdwjidx0C,
 700                                             vdwioffsetptr0+vdwjidx0D,
 701                                             &c6_00,&c12_00);
 702
 703             /* EWALD ELECTROSTATICS */
 704
 705             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 706             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 707             ewitab           = _mm256_cvttpd_epi32(ewrt);
 708             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 709             gmx_mm256_load_4pair_swizzle_pd(ewtab + _mm_extract_epi32(ewitab,0),ewtab + _mm_extract_epi32(ewitab,1),
 710                                             ewtab + _mm_extract_epi32(ewitab,2),ewtab + _mm_extract_epi32(ewitab,3),
 711                                             &ewtabF,&ewtabFn);
 712             felec            = _mm256_add_pd(_mm256_mul_pd( _mm256_sub_pd(one,eweps),ewtabF),_mm256_mul_pd(eweps,ewtabFn));
 713             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 714
 715             /* LENNARD-JONES DISPERSION/REPULSION */
 716
 717             rinvsix          = _mm256_mul_pd(_mm256_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 718             fvdw             = _mm256_mul_pd(_mm256_sub_pd(_mm256_mul_pd(c12_00,rinvsix),c6_00),_mm256_mul_pd(rinvsix,rinvsq00));
 719
 720             fscal            = _mm256_add_pd(felec,fvdw);
 721
 722             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 723
 724             /* Calculate temporary vectorial force */
 725             tx               = _mm256_mul_pd(fscal,dx00);
 726             ty               = _mm256_mul_pd(fscal,dy00);
 727             tz               = _mm256_mul_pd(fscal,dz00);
 728
 729             /* Update vectorial force */
 730             fix0             = _mm256_add_pd(fix0,tx);
 731             fiy0             = _mm256_add_pd(fiy0,ty);
 732             fiz0             = _mm256_add_pd(fiz0,tz);
 733
 734             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 735             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 736             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 737             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 738             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 739
 740             /* Inner loop uses 44 flops */
 741         }
 742
 743         /* End of innermost loop */
 744
 745         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 746                                                  f+i_coord_offset,fshift+i_shift_offset);
 747
 748         /* Increment number of inner iterations */
 749         inneriter                  += j_index_end - j_index_start;
 750
 751         /* Outer loop uses 7 flops */
 752     }
 753
 754     /* Increment number of outer iterations */
 755     outeriter        += nri;
 756
 757     /* Update outer/inner flops */
 758
 759     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*44);
 760 }