src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecEwSw_VdwNone_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 "gromacs/simd/math_x86_avx_256_double.h"
  50 #include "kernelutil_x86_avx_256_double.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwNone_GeomP1P1_VF_avx_256_double
  54  * Electrostatics interaction: Ewald
  55  * VdW interaction:            None
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecEwSw_VdwNone_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     __m128i          ewitab;
  94     __m256d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  95     __m256d          beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
  96     real             *ewtab;
  97     __m256d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  98     real             rswitch_scalar,d_scalar;
  99     __m256d          dummy_mask,cutoff_mask;
 100     __m128           tmpmask0,tmpmask1;
 101     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 102     __m256d          one     = _mm256_set1_pd(1.0);
 103     __m256d          two     = _mm256_set1_pd(2.0);
 104     x                = xx[0];
 105     f                = ff[0];
 106
 107     nri              = nlist->nri;
 108     iinr             = nlist->iinr;
 109     jindex           = nlist->jindex;
 110     jjnr             = nlist->jjnr;
 111     shiftidx         = nlist->shift;
 112     gid              = nlist->gid;
 113     shiftvec         = fr->shift_vec[0];
 114     fshift           = fr->fshift[0];
 115     facel            = _mm256_set1_pd(fr->epsfac);
 116     charge           = mdatoms->chargeA;
 117
 118     sh_ewald         = _mm256_set1_pd(fr->ic->sh_ewald);
 119     beta             = _mm256_set1_pd(fr->ic->ewaldcoeff_q);
 120     beta2            = _mm256_mul_pd(beta,beta);
 121     beta3            = _mm256_mul_pd(beta,beta2);
 122
 123     ewtab            = fr->ic->tabq_coul_FDV0;
 124     ewtabscale       = _mm256_set1_pd(fr->ic->tabq_scale);
 125     ewtabhalfspace   = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
 126
 127     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 128     rcutoff_scalar   = fr->rcoulomb;
 129     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 130     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 131
 132     rswitch_scalar   = fr->rcoulomb_switch;
 133     rswitch          = _mm256_set1_pd(rswitch_scalar);
 134     /* Setup switch parameters */
 135     d_scalar         = rcutoff_scalar-rswitch_scalar;
 136     d                = _mm256_set1_pd(d_scalar);
 137     swV3             = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 138     swV4             = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 139     swV5             = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 140     swF2             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 141     swF3             = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 142     swF4             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 143
 144     /* Avoid stupid compiler warnings */
 145     jnrA = jnrB = jnrC = jnrD = 0;
 146     j_coord_offsetA = 0;
 147     j_coord_offsetB = 0;
 148     j_coord_offsetC = 0;
 149     j_coord_offsetD = 0;
 150
 151     outeriter        = 0;
 152     inneriter        = 0;
 153
 154     for(iidx=0;iidx<4*DIM;iidx++)
 155     {
 156         scratch[iidx] = 0.0;
 157     }
 158
 159     /* Start outer loop over neighborlists */
 160     for(iidx=0; iidx<nri; iidx++)
 161     {
 162         /* Load shift vector for this list */
 163         i_shift_offset   = DIM*shiftidx[iidx];
 164
 165         /* Load limits for loop over neighbors */
 166         j_index_start    = jindex[iidx];
 167         j_index_end      = jindex[iidx+1];
 168
 169         /* Get outer coordinate index */
 170         inr              = iinr[iidx];
 171         i_coord_offset   = DIM*inr;
 172
 173         /* Load i particle coords and add shift vector */
 174         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 175
 176         fix0             = _mm256_setzero_pd();
 177         fiy0             = _mm256_setzero_pd();
 178         fiz0             = _mm256_setzero_pd();
 179
 180         /* Load parameters for i particles */
 181         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 182
 183         /* Reset potential sums */
 184         velecsum         = _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
 221             /**************************
 222              * CALCULATE INTERACTIONS *
 223              **************************/
 224
 225             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 226             {
 227
 228             r00              = _mm256_mul_pd(rsq00,rinv00);
 229
 230             /* Compute parameters for interactions between i and j atoms */
 231             qq00             = _mm256_mul_pd(iq0,jq0);
 232
 233             /* EWALD ELECTROSTATICS */
 234
 235             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 236             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 237             ewitab           = _mm256_cvttpd_epi32(ewrt);
 238             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 239             ewitab           = _mm_slli_epi32(ewitab,2);
 240             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 241             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 242             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 243             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 244             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 245             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 246             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 247             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
 248             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 249
 250             d                = _mm256_sub_pd(r00,rswitch);
 251             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 252             d2               = _mm256_mul_pd(d,d);
 253             sw               = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
 254
 255             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 256
 257             /* Evaluate switch function */
 258             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 259             felec            = _mm256_sub_pd( _mm256_mul_pd(felec,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(velec,dsw)) );
 260             velec            = _mm256_mul_pd(velec,sw);
 261             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 262
 263             /* Update potential sum for this i atom from the interaction with this j atom. */
 264             velec            = _mm256_and_pd(velec,cutoff_mask);
 265             velecsum         = _mm256_add_pd(velecsum,velec);
 266
 267             fscal            = felec;
 268
 269             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 270
 271             /* Calculate temporary vectorial force */
 272             tx               = _mm256_mul_pd(fscal,dx00);
 273             ty               = _mm256_mul_pd(fscal,dy00);
 274             tz               = _mm256_mul_pd(fscal,dz00);
 275
 276             /* Update vectorial force */
 277             fix0             = _mm256_add_pd(fix0,tx);
 278             fiy0             = _mm256_add_pd(fiy0,ty);
 279             fiz0             = _mm256_add_pd(fiz0,tz);
 280
 281             fjptrA             = f+j_coord_offsetA;
 282             fjptrB             = f+j_coord_offsetB;
 283             fjptrC             = f+j_coord_offsetC;
 284             fjptrD             = f+j_coord_offsetD;
 285             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 286
 287             }
 288
 289             /* Inner loop uses 65 flops */
 290         }
 291
 292         if(jidx<j_index_end)
 293         {
 294
 295             /* Get j neighbor index, and coordinate index */
 296             jnrlistA         = jjnr[jidx];
 297             jnrlistB         = jjnr[jidx+1];
 298             jnrlistC         = jjnr[jidx+2];
 299             jnrlistD         = jjnr[jidx+3];
 300             /* Sign of each element will be negative for non-real atoms.
 301              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 302              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 303              */
 304             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 305
 306             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 307             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 308             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 309
 310             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 311             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 312             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 313             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 314             j_coord_offsetA  = DIM*jnrA;
 315             j_coord_offsetB  = DIM*jnrB;
 316             j_coord_offsetC  = DIM*jnrC;
 317             j_coord_offsetD  = DIM*jnrD;
 318
 319             /* load j atom coordinates */
 320             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 321                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 322                                                  &jx0,&jy0,&jz0);
 323
 324             /* Calculate displacement vector */
 325             dx00             = _mm256_sub_pd(ix0,jx0);
 326             dy00             = _mm256_sub_pd(iy0,jy0);
 327             dz00             = _mm256_sub_pd(iz0,jz0);
 328
 329             /* Calculate squared distance and things based on it */
 330             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 331
 332             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 333
 334             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 335
 336             /* Load parameters for j particles */
 337             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 338                                                                  charge+jnrC+0,charge+jnrD+0);
 339
 340             /**************************
 341              * CALCULATE INTERACTIONS *
 342              **************************/
 343
 344             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 345             {
 346
 347             r00              = _mm256_mul_pd(rsq00,rinv00);
 348             r00              = _mm256_andnot_pd(dummy_mask,r00);
 349
 350             /* Compute parameters for interactions between i and j atoms */
 351             qq00             = _mm256_mul_pd(iq0,jq0);
 352
 353             /* EWALD ELECTROSTATICS */
 354
 355             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 356             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 357             ewitab           = _mm256_cvttpd_epi32(ewrt);
 358             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 359             ewitab           = _mm_slli_epi32(ewitab,2);
 360             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 361             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 362             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 363             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 364             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 365             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 366             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 367             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
 368             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 369
 370             d                = _mm256_sub_pd(r00,rswitch);
 371             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 372             d2               = _mm256_mul_pd(d,d);
 373             sw               = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
 374
 375             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 376
 377             /* Evaluate switch function */
 378             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 379             felec            = _mm256_sub_pd( _mm256_mul_pd(felec,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(velec,dsw)) );
 380             velec            = _mm256_mul_pd(velec,sw);
 381             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 382
 383             /* Update potential sum for this i atom from the interaction with this j atom. */
 384             velec            = _mm256_and_pd(velec,cutoff_mask);
 385             velec            = _mm256_andnot_pd(dummy_mask,velec);
 386             velecsum         = _mm256_add_pd(velecsum,velec);
 387
 388             fscal            = felec;
 389
 390             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 391
 392             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 393
 394             /* Calculate temporary vectorial force */
 395             tx               = _mm256_mul_pd(fscal,dx00);
 396             ty               = _mm256_mul_pd(fscal,dy00);
 397             tz               = _mm256_mul_pd(fscal,dz00);
 398
 399             /* Update vectorial force */
 400             fix0             = _mm256_add_pd(fix0,tx);
 401             fiy0             = _mm256_add_pd(fiy0,ty);
 402             fiz0             = _mm256_add_pd(fiz0,tz);
 403
 404             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 405             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 406             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 407             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 408             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 409
 410             }
 411
 412             /* Inner loop uses 66 flops */
 413         }
 414
 415         /* End of innermost loop */
 416
 417         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 418                                                  f+i_coord_offset,fshift+i_shift_offset);
 419
 420         ggid                        = gid[iidx];
 421         /* Update potential energies */
 422         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 423
 424         /* Increment number of inner iterations */
 425         inneriter                  += j_index_end - j_index_start;
 426
 427         /* Outer loop uses 8 flops */
 428     }
 429
 430     /* Increment number of outer iterations */
 431     outeriter        += nri;
 432
 433     /* Update outer/inner flops */
 434
 435     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*66);
 436 }
 437 /*
 438  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwNone_GeomP1P1_F_avx_256_double
 439  * Electrostatics interaction: Ewald
 440  * VdW interaction:            None
 441  * Geometry:                   Particle-Particle
 442  * Calculate force/pot:        Force
 443  */
 444 void
 445 nb_kernel_ElecEwSw_VdwNone_GeomP1P1_F_avx_256_double
 446                     (t_nblist                    * gmx_restrict       nlist,
 447                      rvec                        * gmx_restrict          xx,
 448                      rvec                        * gmx_restrict          ff,
 449                      t_forcerec                  * gmx_restrict          fr,
 450                      t_mdatoms                   * gmx_restrict     mdatoms,
 451                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 452                      t_nrnb                      * gmx_restrict        nrnb)
 453 {
 454     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 455      * just 0 for non-waters.
 456      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 457      * jnr indices corresponding to data put in the four positions in the SIMD register.
 458      */
 459     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 460     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 461     int              jnrA,jnrB,jnrC,jnrD;
 462     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 463     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 464     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 465     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 466     real             rcutoff_scalar;
 467     real             *shiftvec,*fshift,*x,*f;
 468     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 469     real             scratch[4*DIM];
 470     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 471     real *           vdwioffsetptr0;
 472     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 473     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 474     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 475     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 476     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 477     real             *charge;
 478     __m128i          ewitab;
 479     __m256d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 480     __m256d          beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 481     real             *ewtab;
 482     __m256d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 483     real             rswitch_scalar,d_scalar;
 484     __m256d          dummy_mask,cutoff_mask;
 485     __m128           tmpmask0,tmpmask1;
 486     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 487     __m256d          one     = _mm256_set1_pd(1.0);
 488     __m256d          two     = _mm256_set1_pd(2.0);
 489     x                = xx[0];
 490     f                = ff[0];
 491
 492     nri              = nlist->nri;
 493     iinr             = nlist->iinr;
 494     jindex           = nlist->jindex;
 495     jjnr             = nlist->jjnr;
 496     shiftidx         = nlist->shift;
 497     gid              = nlist->gid;
 498     shiftvec         = fr->shift_vec[0];
 499     fshift           = fr->fshift[0];
 500     facel            = _mm256_set1_pd(fr->epsfac);
 501     charge           = mdatoms->chargeA;
 502
 503     sh_ewald         = _mm256_set1_pd(fr->ic->sh_ewald);
 504     beta             = _mm256_set1_pd(fr->ic->ewaldcoeff_q);
 505     beta2            = _mm256_mul_pd(beta,beta);
 506     beta3            = _mm256_mul_pd(beta,beta2);
 507
 508     ewtab            = fr->ic->tabq_coul_FDV0;
 509     ewtabscale       = _mm256_set1_pd(fr->ic->tabq_scale);
 510     ewtabhalfspace   = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
 511
 512     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 513     rcutoff_scalar   = fr->rcoulomb;
 514     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 515     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 516
 517     rswitch_scalar   = fr->rcoulomb_switch;
 518     rswitch          = _mm256_set1_pd(rswitch_scalar);
 519     /* Setup switch parameters */
 520     d_scalar         = rcutoff_scalar-rswitch_scalar;
 521     d                = _mm256_set1_pd(d_scalar);
 522     swV3             = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 523     swV4             = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 524     swV5             = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 525     swF2             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 526     swF3             = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 527     swF4             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 528
 529     /* Avoid stupid compiler warnings */
 530     jnrA = jnrB = jnrC = jnrD = 0;
 531     j_coord_offsetA = 0;
 532     j_coord_offsetB = 0;
 533     j_coord_offsetC = 0;
 534     j_coord_offsetD = 0;
 535
 536     outeriter        = 0;
 537     inneriter        = 0;
 538
 539     for(iidx=0;iidx<4*DIM;iidx++)
 540     {
 541         scratch[iidx] = 0.0;
 542     }
 543
 544     /* Start outer loop over neighborlists */
 545     for(iidx=0; iidx<nri; iidx++)
 546     {
 547         /* Load shift vector for this list */
 548         i_shift_offset   = DIM*shiftidx[iidx];
 549
 550         /* Load limits for loop over neighbors */
 551         j_index_start    = jindex[iidx];
 552         j_index_end      = jindex[iidx+1];
 553
 554         /* Get outer coordinate index */
 555         inr              = iinr[iidx];
 556         i_coord_offset   = DIM*inr;
 557
 558         /* Load i particle coords and add shift vector */
 559         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 560
 561         fix0             = _mm256_setzero_pd();
 562         fiy0             = _mm256_setzero_pd();
 563         fiz0             = _mm256_setzero_pd();
 564
 565         /* Load parameters for i particles */
 566         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 567
 568         /* Start inner kernel loop */
 569         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 570         {
 571
 572             /* Get j neighbor index, and coordinate index */
 573             jnrA             = jjnr[jidx];
 574             jnrB             = jjnr[jidx+1];
 575             jnrC             = jjnr[jidx+2];
 576             jnrD             = jjnr[jidx+3];
 577             j_coord_offsetA  = DIM*jnrA;
 578             j_coord_offsetB  = DIM*jnrB;
 579             j_coord_offsetC  = DIM*jnrC;
 580             j_coord_offsetD  = DIM*jnrD;
 581
 582             /* load j atom coordinates */
 583             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 584                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 585                                                  &jx0,&jy0,&jz0);
 586
 587             /* Calculate displacement vector */
 588             dx00             = _mm256_sub_pd(ix0,jx0);
 589             dy00             = _mm256_sub_pd(iy0,jy0);
 590             dz00             = _mm256_sub_pd(iz0,jz0);
 591
 592             /* Calculate squared distance and things based on it */
 593             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 594
 595             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 596
 597             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 598
 599             /* Load parameters for j particles */
 600             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 601                                                                  charge+jnrC+0,charge+jnrD+0);
 602
 603             /**************************
 604              * CALCULATE INTERACTIONS *
 605              **************************/
 606
 607             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 608             {
 609
 610             r00              = _mm256_mul_pd(rsq00,rinv00);
 611
 612             /* Compute parameters for interactions between i and j atoms */
 613             qq00             = _mm256_mul_pd(iq0,jq0);
 614
 615             /* EWALD ELECTROSTATICS */
 616
 617             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 618             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 619             ewitab           = _mm256_cvttpd_epi32(ewrt);
 620             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 621             ewitab           = _mm_slli_epi32(ewitab,2);
 622             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 623             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 624             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 625             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 626             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 627             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 628             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 629             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
 630             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 631
 632             d                = _mm256_sub_pd(r00,rswitch);
 633             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 634             d2               = _mm256_mul_pd(d,d);
 635             sw               = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
 636
 637             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 638
 639             /* Evaluate switch function */
 640             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 641             felec            = _mm256_sub_pd( _mm256_mul_pd(felec,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(velec,dsw)) );
 642             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 643
 644             fscal            = felec;
 645
 646             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 647
 648             /* Calculate temporary vectorial force */
 649             tx               = _mm256_mul_pd(fscal,dx00);
 650             ty               = _mm256_mul_pd(fscal,dy00);
 651             tz               = _mm256_mul_pd(fscal,dz00);
 652
 653             /* Update vectorial force */
 654             fix0             = _mm256_add_pd(fix0,tx);
 655             fiy0             = _mm256_add_pd(fiy0,ty);
 656             fiz0             = _mm256_add_pd(fiz0,tz);
 657
 658             fjptrA             = f+j_coord_offsetA;
 659             fjptrB             = f+j_coord_offsetB;
 660             fjptrC             = f+j_coord_offsetC;
 661             fjptrD             = f+j_coord_offsetD;
 662             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 663
 664             }
 665
 666             /* Inner loop uses 62 flops */
 667         }
 668
 669         if(jidx<j_index_end)
 670         {
 671
 672             /* Get j neighbor index, and coordinate index */
 673             jnrlistA         = jjnr[jidx];
 674             jnrlistB         = jjnr[jidx+1];
 675             jnrlistC         = jjnr[jidx+2];
 676             jnrlistD         = jjnr[jidx+3];
 677             /* Sign of each element will be negative for non-real atoms.
 678              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 679              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 680              */
 681             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 682
 683             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 684             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 685             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 686
 687             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 688             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 689             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 690             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 691             j_coord_offsetA  = DIM*jnrA;
 692             j_coord_offsetB  = DIM*jnrB;
 693             j_coord_offsetC  = DIM*jnrC;
 694             j_coord_offsetD  = DIM*jnrD;
 695
 696             /* load j atom coordinates */
 697             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 698                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 699                                                  &jx0,&jy0,&jz0);
 700
 701             /* Calculate displacement vector */
 702             dx00             = _mm256_sub_pd(ix0,jx0);
 703             dy00             = _mm256_sub_pd(iy0,jy0);
 704             dz00             = _mm256_sub_pd(iz0,jz0);
 705
 706             /* Calculate squared distance and things based on it */
 707             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 708
 709             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 710
 711             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 712
 713             /* Load parameters for j particles */
 714             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 715                                                                  charge+jnrC+0,charge+jnrD+0);
 716
 717             /**************************
 718              * CALCULATE INTERACTIONS *
 719              **************************/
 720
 721             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 722             {
 723
 724             r00              = _mm256_mul_pd(rsq00,rinv00);
 725             r00              = _mm256_andnot_pd(dummy_mask,r00);
 726
 727             /* Compute parameters for interactions between i and j atoms */
 728             qq00             = _mm256_mul_pd(iq0,jq0);
 729
 730             /* EWALD ELECTROSTATICS */
 731
 732             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 733             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 734             ewitab           = _mm256_cvttpd_epi32(ewrt);
 735             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 736             ewitab           = _mm_slli_epi32(ewitab,2);
 737             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 738             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 739             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 740             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 741             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 742             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 743             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 744             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
 745             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 746
 747             d                = _mm256_sub_pd(r00,rswitch);
 748             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 749             d2               = _mm256_mul_pd(d,d);
 750             sw               = _mm256_add_pd(one,_mm256_mul_pd(d2,_mm256_mul_pd(d,_mm256_add_pd(swV3,_mm256_mul_pd(d,_mm256_add_pd(swV4,_mm256_mul_pd(d,swV5)))))));
 751
 752             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 753
 754             /* Evaluate switch function */
 755             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 756             felec            = _mm256_sub_pd( _mm256_mul_pd(felec,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(velec,dsw)) );
 757             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 758
 759             fscal            = felec;
 760
 761             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 762
 763             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 764
 765             /* Calculate temporary vectorial force */
 766             tx               = _mm256_mul_pd(fscal,dx00);
 767             ty               = _mm256_mul_pd(fscal,dy00);
 768             tz               = _mm256_mul_pd(fscal,dz00);
 769
 770             /* Update vectorial force */
 771             fix0             = _mm256_add_pd(fix0,tx);
 772             fiy0             = _mm256_add_pd(fiy0,ty);
 773             fiz0             = _mm256_add_pd(fiz0,tz);
 774
 775             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 776             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 777             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 778             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 779             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 780
 781             }
 782
 783             /* Inner loop uses 63 flops */
 784         }
 785
 786         /* End of innermost loop */
 787
 788         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 789                                                  f+i_coord_offset,fshift+i_shift_offset);
 790
 791         /* Increment number of inner iterations */
 792         inneriter                  += j_index_end - j_index_start;
 793
 794         /* Outer loop uses 7 flops */
 795     }
 796
 797     /* Increment number of outer iterations */
 798     outeriter        += nri;
 799
 800     /* Update outer/inner flops */
 801
 802     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*63);
 803 }