src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_single/nb_kernel_ElecEwSw_VdwNone_GeomP1P1_avx_256_single.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, 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_single kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_256_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwNone_GeomP1P1_VF_avx_256_single
  51  * Electrostatics interaction: Ewald
  52  * VdW interaction:            None
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecEwSw_VdwNone_GeomP1P1_VF_avx_256_single
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrE,jnrF,jnrG,jnrH;
  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              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
  83     real             scratch[4*DIM];
  84     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  85     real *           vdwioffsetptr0;
  86     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
  88     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  90     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     __m256i          ewitab;
  93     __m128i          ewitab_lo,ewitab_hi;
  94     __m256           ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  95     __m256           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
  96     real             *ewtab;
  97     __m256           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  98     real             rswitch_scalar,d_scalar;
  99     __m256           dummy_mask,cutoff_mask;
 100     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 101     __m256           one     = _mm256_set1_ps(1.0);
 102     __m256           two     = _mm256_set1_ps(2.0);
 103     x                = xx[0];
 104     f                = ff[0];
 105
 106     nri              = nlist->nri;
 107     iinr             = nlist->iinr;
 108     jindex           = nlist->jindex;
 109     jjnr             = nlist->jjnr;
 110     shiftidx         = nlist->shift;
 111     gid              = nlist->gid;
 112     shiftvec         = fr->shift_vec[0];
 113     fshift           = fr->fshift[0];
 114     facel            = _mm256_set1_ps(fr->ic->epsfac);
 115     charge           = mdatoms->chargeA;
 116
 117     sh_ewald         = _mm256_set1_ps(fr->ic->sh_ewald);
 118     beta             = _mm256_set1_ps(fr->ic->ewaldcoeff_q);
 119     beta2            = _mm256_mul_ps(beta,beta);
 120     beta3            = _mm256_mul_ps(beta,beta2);
 121
 122     ewtab            = fr->ic->tabq_coul_FDV0;
 123     ewtabscale       = _mm256_set1_ps(fr->ic->tabq_scale);
 124     ewtabhalfspace   = _mm256_set1_ps(0.5/fr->ic->tabq_scale);
 125
 126     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 127     rcutoff_scalar   = fr->ic->rcoulomb;
 128     rcutoff          = _mm256_set1_ps(rcutoff_scalar);
 129     rcutoff2         = _mm256_mul_ps(rcutoff,rcutoff);
 130
 131     rswitch_scalar   = fr->ic->rcoulomb_switch;
 132     rswitch          = _mm256_set1_ps(rswitch_scalar);
 133     /* Setup switch parameters */
 134     d_scalar         = rcutoff_scalar-rswitch_scalar;
 135     d                = _mm256_set1_ps(d_scalar);
 136     swV3             = _mm256_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 137     swV4             = _mm256_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 138     swV5             = _mm256_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 139     swF2             = _mm256_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 140     swF3             = _mm256_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 141     swF4             = _mm256_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 142
 143     /* Avoid stupid compiler warnings */
 144     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 145     j_coord_offsetA = 0;
 146     j_coord_offsetB = 0;
 147     j_coord_offsetC = 0;
 148     j_coord_offsetD = 0;
 149     j_coord_offsetE = 0;
 150     j_coord_offsetF = 0;
 151     j_coord_offsetG = 0;
 152     j_coord_offsetH = 0;
 153
 154     outeriter        = 0;
 155     inneriter        = 0;
 156
 157     for(iidx=0;iidx<4*DIM;iidx++)
 158     {
 159         scratch[iidx] = 0.0;
 160     }
 161
 162     /* Start outer loop over neighborlists */
 163     for(iidx=0; iidx<nri; iidx++)
 164     {
 165         /* Load shift vector for this list */
 166         i_shift_offset   = DIM*shiftidx[iidx];
 167
 168         /* Load limits for loop over neighbors */
 169         j_index_start    = jindex[iidx];
 170         j_index_end      = jindex[iidx+1];
 171
 172         /* Get outer coordinate index */
 173         inr              = iinr[iidx];
 174         i_coord_offset   = DIM*inr;
 175
 176         /* Load i particle coords and add shift vector */
 177         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 178
 179         fix0             = _mm256_setzero_ps();
 180         fiy0             = _mm256_setzero_ps();
 181         fiz0             = _mm256_setzero_ps();
 182
 183         /* Load parameters for i particles */
 184         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 185
 186         /* Reset potential sums */
 187         velecsum         = _mm256_setzero_ps();
 188
 189         /* Start inner kernel loop */
 190         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 191         {
 192
 193             /* Get j neighbor index, and coordinate index */
 194             jnrA             = jjnr[jidx];
 195             jnrB             = jjnr[jidx+1];
 196             jnrC             = jjnr[jidx+2];
 197             jnrD             = jjnr[jidx+3];
 198             jnrE             = jjnr[jidx+4];
 199             jnrF             = jjnr[jidx+5];
 200             jnrG             = jjnr[jidx+6];
 201             jnrH             = jjnr[jidx+7];
 202             j_coord_offsetA  = DIM*jnrA;
 203             j_coord_offsetB  = DIM*jnrB;
 204             j_coord_offsetC  = DIM*jnrC;
 205             j_coord_offsetD  = DIM*jnrD;
 206             j_coord_offsetE  = DIM*jnrE;
 207             j_coord_offsetF  = DIM*jnrF;
 208             j_coord_offsetG  = DIM*jnrG;
 209             j_coord_offsetH  = DIM*jnrH;
 210
 211             /* load j atom coordinates */
 212             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 213                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 214                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 215                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 216                                                  &jx0,&jy0,&jz0);
 217
 218             /* Calculate displacement vector */
 219             dx00             = _mm256_sub_ps(ix0,jx0);
 220             dy00             = _mm256_sub_ps(iy0,jy0);
 221             dz00             = _mm256_sub_ps(iz0,jz0);
 222
 223             /* Calculate squared distance and things based on it */
 224             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 225
 226             rinv00           = avx256_invsqrt_f(rsq00);
 227
 228             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 229
 230             /* Load parameters for j particles */
 231             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 232                                                                  charge+jnrC+0,charge+jnrD+0,
 233                                                                  charge+jnrE+0,charge+jnrF+0,
 234                                                                  charge+jnrG+0,charge+jnrH+0);
 235
 236             /**************************
 237              * CALCULATE INTERACTIONS *
 238              **************************/
 239
 240             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 241             {
 242
 243             r00              = _mm256_mul_ps(rsq00,rinv00);
 244
 245             /* Compute parameters for interactions between i and j atoms */
 246             qq00             = _mm256_mul_ps(iq0,jq0);
 247
 248             /* EWALD ELECTROSTATICS */
 249
 250             /* Analytical PME correction */
 251             zeta2            = _mm256_mul_ps(beta2,rsq00);
 252             rinv3            = _mm256_mul_ps(rinvsq00,rinv00);
 253             pmecorrF         = avx256_pmecorrF_f(zeta2);
 254             felec            = _mm256_add_ps( _mm256_mul_ps(pmecorrF,beta3), rinv3);
 255             felec            = _mm256_mul_ps(qq00,felec);
 256             pmecorrV         = avx256_pmecorrV_f(zeta2);
 257             pmecorrV         = _mm256_mul_ps(pmecorrV,beta);
 258             velec            = _mm256_sub_ps(rinv00,pmecorrV);
 259             velec            = _mm256_mul_ps(qq00,velec);
 260
 261             d                = _mm256_sub_ps(r00,rswitch);
 262             d                = _mm256_max_ps(d,_mm256_setzero_ps());
 263             d2               = _mm256_mul_ps(d,d);
 264             sw               = _mm256_add_ps(one,_mm256_mul_ps(d2,_mm256_mul_ps(d,_mm256_add_ps(swV3,_mm256_mul_ps(d,_mm256_add_ps(swV4,_mm256_mul_ps(d,swV5)))))));
 265
 266             dsw              = _mm256_mul_ps(d2,_mm256_add_ps(swF2,_mm256_mul_ps(d,_mm256_add_ps(swF3,_mm256_mul_ps(d,swF4)))));
 267
 268             /* Evaluate switch function */
 269             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 270             felec            = _mm256_sub_ps( _mm256_mul_ps(felec,sw) , _mm256_mul_ps(rinv00,_mm256_mul_ps(velec,dsw)) );
 271             velec            = _mm256_mul_ps(velec,sw);
 272             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 273
 274             /* Update potential sum for this i atom from the interaction with this j atom. */
 275             velec            = _mm256_and_ps(velec,cutoff_mask);
 276             velecsum         = _mm256_add_ps(velecsum,velec);
 277
 278             fscal            = felec;
 279
 280             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 281
 282             /* Calculate temporary vectorial force */
 283             tx               = _mm256_mul_ps(fscal,dx00);
 284             ty               = _mm256_mul_ps(fscal,dy00);
 285             tz               = _mm256_mul_ps(fscal,dz00);
 286
 287             /* Update vectorial force */
 288             fix0             = _mm256_add_ps(fix0,tx);
 289             fiy0             = _mm256_add_ps(fiy0,ty);
 290             fiz0             = _mm256_add_ps(fiz0,tz);
 291
 292             fjptrA             = f+j_coord_offsetA;
 293             fjptrB             = f+j_coord_offsetB;
 294             fjptrC             = f+j_coord_offsetC;
 295             fjptrD             = f+j_coord_offsetD;
 296             fjptrE             = f+j_coord_offsetE;
 297             fjptrF             = f+j_coord_offsetF;
 298             fjptrG             = f+j_coord_offsetG;
 299             fjptrH             = f+j_coord_offsetH;
 300             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 301
 302             }
 303
 304             /* Inner loop uses 108 flops */
 305         }
 306
 307         if(jidx<j_index_end)
 308         {
 309
 310             /* Get j neighbor index, and coordinate index */
 311             jnrlistA         = jjnr[jidx];
 312             jnrlistB         = jjnr[jidx+1];
 313             jnrlistC         = jjnr[jidx+2];
 314             jnrlistD         = jjnr[jidx+3];
 315             jnrlistE         = jjnr[jidx+4];
 316             jnrlistF         = jjnr[jidx+5];
 317             jnrlistG         = jjnr[jidx+6];
 318             jnrlistH         = jjnr[jidx+7];
 319             /* Sign of each element will be negative for non-real atoms.
 320              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 321              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 322              */
 323             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 324                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 325
 326             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 327             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 328             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 329             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 330             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 331             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 332             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 333             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 334             j_coord_offsetA  = DIM*jnrA;
 335             j_coord_offsetB  = DIM*jnrB;
 336             j_coord_offsetC  = DIM*jnrC;
 337             j_coord_offsetD  = DIM*jnrD;
 338             j_coord_offsetE  = DIM*jnrE;
 339             j_coord_offsetF  = DIM*jnrF;
 340             j_coord_offsetG  = DIM*jnrG;
 341             j_coord_offsetH  = DIM*jnrH;
 342
 343             /* load j atom coordinates */
 344             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 345                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 346                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 347                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 348                                                  &jx0,&jy0,&jz0);
 349
 350             /* Calculate displacement vector */
 351             dx00             = _mm256_sub_ps(ix0,jx0);
 352             dy00             = _mm256_sub_ps(iy0,jy0);
 353             dz00             = _mm256_sub_ps(iz0,jz0);
 354
 355             /* Calculate squared distance and things based on it */
 356             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 357
 358             rinv00           = avx256_invsqrt_f(rsq00);
 359
 360             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 361
 362             /* Load parameters for j particles */
 363             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 364                                                                  charge+jnrC+0,charge+jnrD+0,
 365                                                                  charge+jnrE+0,charge+jnrF+0,
 366                                                                  charge+jnrG+0,charge+jnrH+0);
 367
 368             /**************************
 369              * CALCULATE INTERACTIONS *
 370              **************************/
 371
 372             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 373             {
 374
 375             r00              = _mm256_mul_ps(rsq00,rinv00);
 376             r00              = _mm256_andnot_ps(dummy_mask,r00);
 377
 378             /* Compute parameters for interactions between i and j atoms */
 379             qq00             = _mm256_mul_ps(iq0,jq0);
 380
 381             /* EWALD ELECTROSTATICS */
 382
 383             /* Analytical PME correction */
 384             zeta2            = _mm256_mul_ps(beta2,rsq00);
 385             rinv3            = _mm256_mul_ps(rinvsq00,rinv00);
 386             pmecorrF         = avx256_pmecorrF_f(zeta2);
 387             felec            = _mm256_add_ps( _mm256_mul_ps(pmecorrF,beta3), rinv3);
 388             felec            = _mm256_mul_ps(qq00,felec);
 389             pmecorrV         = avx256_pmecorrV_f(zeta2);
 390             pmecorrV         = _mm256_mul_ps(pmecorrV,beta);
 391             velec            = _mm256_sub_ps(rinv00,pmecorrV);
 392             velec            = _mm256_mul_ps(qq00,velec);
 393
 394             d                = _mm256_sub_ps(r00,rswitch);
 395             d                = _mm256_max_ps(d,_mm256_setzero_ps());
 396             d2               = _mm256_mul_ps(d,d);
 397             sw               = _mm256_add_ps(one,_mm256_mul_ps(d2,_mm256_mul_ps(d,_mm256_add_ps(swV3,_mm256_mul_ps(d,_mm256_add_ps(swV4,_mm256_mul_ps(d,swV5)))))));
 398
 399             dsw              = _mm256_mul_ps(d2,_mm256_add_ps(swF2,_mm256_mul_ps(d,_mm256_add_ps(swF3,_mm256_mul_ps(d,swF4)))));
 400
 401             /* Evaluate switch function */
 402             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 403             felec            = _mm256_sub_ps( _mm256_mul_ps(felec,sw) , _mm256_mul_ps(rinv00,_mm256_mul_ps(velec,dsw)) );
 404             velec            = _mm256_mul_ps(velec,sw);
 405             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 406
 407             /* Update potential sum for this i atom from the interaction with this j atom. */
 408             velec            = _mm256_and_ps(velec,cutoff_mask);
 409             velec            = _mm256_andnot_ps(dummy_mask,velec);
 410             velecsum         = _mm256_add_ps(velecsum,velec);
 411
 412             fscal            = felec;
 413
 414             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 415
 416             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 417
 418             /* Calculate temporary vectorial force */
 419             tx               = _mm256_mul_ps(fscal,dx00);
 420             ty               = _mm256_mul_ps(fscal,dy00);
 421             tz               = _mm256_mul_ps(fscal,dz00);
 422
 423             /* Update vectorial force */
 424             fix0             = _mm256_add_ps(fix0,tx);
 425             fiy0             = _mm256_add_ps(fiy0,ty);
 426             fiz0             = _mm256_add_ps(fiz0,tz);
 427
 428             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 429             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 430             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 431             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 432             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 433             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 434             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 435             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 436             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 437
 438             }
 439
 440             /* Inner loop uses 109 flops */
 441         }
 442
 443         /* End of innermost loop */
 444
 445         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 446                                                  f+i_coord_offset,fshift+i_shift_offset);
 447
 448         ggid                        = gid[iidx];
 449         /* Update potential energies */
 450         gmx_mm256_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 451
 452         /* Increment number of inner iterations */
 453         inneriter                  += j_index_end - j_index_start;
 454
 455         /* Outer loop uses 8 flops */
 456     }
 457
 458     /* Increment number of outer iterations */
 459     outeriter        += nri;
 460
 461     /* Update outer/inner flops */
 462
 463     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*109);
 464 }
 465 /*
 466  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwNone_GeomP1P1_F_avx_256_single
 467  * Electrostatics interaction: Ewald
 468  * VdW interaction:            None
 469  * Geometry:                   Particle-Particle
 470  * Calculate force/pot:        Force
 471  */
 472 void
 473 nb_kernel_ElecEwSw_VdwNone_GeomP1P1_F_avx_256_single
 474                     (t_nblist                    * gmx_restrict       nlist,
 475                      rvec                        * gmx_restrict          xx,
 476                      rvec                        * gmx_restrict          ff,
 477                      struct t_forcerec           * gmx_restrict          fr,
 478                      t_mdatoms                   * gmx_restrict     mdatoms,
 479                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 480                      t_nrnb                      * gmx_restrict        nrnb)
 481 {
 482     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 483      * just 0 for non-waters.
 484      * Suffixes A,B,C,D,E,F,G,H refer to j loop unrolling done with AVX, e.g. for the eight different
 485      * jnr indices corresponding to data put in the four positions in the SIMD register.
 486      */
 487     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 488     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 489     int              jnrA,jnrB,jnrC,jnrD;
 490     int              jnrE,jnrF,jnrG,jnrH;
 491     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 492     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 493     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 494     int              j_coord_offsetE,j_coord_offsetF,j_coord_offsetG,j_coord_offsetH;
 495     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 496     real             rcutoff_scalar;
 497     real             *shiftvec,*fshift,*x,*f;
 498     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD,*fjptrE,*fjptrF,*fjptrG,*fjptrH;
 499     real             scratch[4*DIM];
 500     __m256           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 501     real *           vdwioffsetptr0;
 502     __m256           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 503     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D,vdwjidx0E,vdwjidx0F,vdwjidx0G,vdwjidx0H;
 504     __m256           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 505     __m256           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 506     __m256           velec,felec,velecsum,facel,crf,krf,krf2;
 507     real             *charge;
 508     __m256i          ewitab;
 509     __m128i          ewitab_lo,ewitab_hi;
 510     __m256           ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 511     __m256           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 512     real             *ewtab;
 513     __m256           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 514     real             rswitch_scalar,d_scalar;
 515     __m256           dummy_mask,cutoff_mask;
 516     __m256           signbit = _mm256_castsi256_ps( _mm256_set1_epi32(0x80000000) );
 517     __m256           one     = _mm256_set1_ps(1.0);
 518     __m256           two     = _mm256_set1_ps(2.0);
 519     x                = xx[0];
 520     f                = ff[0];
 521
 522     nri              = nlist->nri;
 523     iinr             = nlist->iinr;
 524     jindex           = nlist->jindex;
 525     jjnr             = nlist->jjnr;
 526     shiftidx         = nlist->shift;
 527     gid              = nlist->gid;
 528     shiftvec         = fr->shift_vec[0];
 529     fshift           = fr->fshift[0];
 530     facel            = _mm256_set1_ps(fr->ic->epsfac);
 531     charge           = mdatoms->chargeA;
 532
 533     sh_ewald         = _mm256_set1_ps(fr->ic->sh_ewald);
 534     beta             = _mm256_set1_ps(fr->ic->ewaldcoeff_q);
 535     beta2            = _mm256_mul_ps(beta,beta);
 536     beta3            = _mm256_mul_ps(beta,beta2);
 537
 538     ewtab            = fr->ic->tabq_coul_FDV0;
 539     ewtabscale       = _mm256_set1_ps(fr->ic->tabq_scale);
 540     ewtabhalfspace   = _mm256_set1_ps(0.5/fr->ic->tabq_scale);
 541
 542     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 543     rcutoff_scalar   = fr->ic->rcoulomb;
 544     rcutoff          = _mm256_set1_ps(rcutoff_scalar);
 545     rcutoff2         = _mm256_mul_ps(rcutoff,rcutoff);
 546
 547     rswitch_scalar   = fr->ic->rcoulomb_switch;
 548     rswitch          = _mm256_set1_ps(rswitch_scalar);
 549     /* Setup switch parameters */
 550     d_scalar         = rcutoff_scalar-rswitch_scalar;
 551     d                = _mm256_set1_ps(d_scalar);
 552     swV3             = _mm256_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 553     swV4             = _mm256_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 554     swV5             = _mm256_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 555     swF2             = _mm256_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 556     swF3             = _mm256_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 557     swF4             = _mm256_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 558
 559     /* Avoid stupid compiler warnings */
 560     jnrA = jnrB = jnrC = jnrD = jnrE = jnrF = jnrG = jnrH = 0;
 561     j_coord_offsetA = 0;
 562     j_coord_offsetB = 0;
 563     j_coord_offsetC = 0;
 564     j_coord_offsetD = 0;
 565     j_coord_offsetE = 0;
 566     j_coord_offsetF = 0;
 567     j_coord_offsetG = 0;
 568     j_coord_offsetH = 0;
 569
 570     outeriter        = 0;
 571     inneriter        = 0;
 572
 573     for(iidx=0;iidx<4*DIM;iidx++)
 574     {
 575         scratch[iidx] = 0.0;
 576     }
 577
 578     /* Start outer loop over neighborlists */
 579     for(iidx=0; iidx<nri; iidx++)
 580     {
 581         /* Load shift vector for this list */
 582         i_shift_offset   = DIM*shiftidx[iidx];
 583
 584         /* Load limits for loop over neighbors */
 585         j_index_start    = jindex[iidx];
 586         j_index_end      = jindex[iidx+1];
 587
 588         /* Get outer coordinate index */
 589         inr              = iinr[iidx];
 590         i_coord_offset   = DIM*inr;
 591
 592         /* Load i particle coords and add shift vector */
 593         gmx_mm256_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 594
 595         fix0             = _mm256_setzero_ps();
 596         fiy0             = _mm256_setzero_ps();
 597         fiz0             = _mm256_setzero_ps();
 598
 599         /* Load parameters for i particles */
 600         iq0              = _mm256_mul_ps(facel,_mm256_set1_ps(charge[inr+0]));
 601
 602         /* Start inner kernel loop */
 603         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+7]>=0; jidx+=8)
 604         {
 605
 606             /* Get j neighbor index, and coordinate index */
 607             jnrA             = jjnr[jidx];
 608             jnrB             = jjnr[jidx+1];
 609             jnrC             = jjnr[jidx+2];
 610             jnrD             = jjnr[jidx+3];
 611             jnrE             = jjnr[jidx+4];
 612             jnrF             = jjnr[jidx+5];
 613             jnrG             = jjnr[jidx+6];
 614             jnrH             = jjnr[jidx+7];
 615             j_coord_offsetA  = DIM*jnrA;
 616             j_coord_offsetB  = DIM*jnrB;
 617             j_coord_offsetC  = DIM*jnrC;
 618             j_coord_offsetD  = DIM*jnrD;
 619             j_coord_offsetE  = DIM*jnrE;
 620             j_coord_offsetF  = DIM*jnrF;
 621             j_coord_offsetG  = DIM*jnrG;
 622             j_coord_offsetH  = DIM*jnrH;
 623
 624             /* load j atom coordinates */
 625             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 626                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 627                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 628                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 629                                                  &jx0,&jy0,&jz0);
 630
 631             /* Calculate displacement vector */
 632             dx00             = _mm256_sub_ps(ix0,jx0);
 633             dy00             = _mm256_sub_ps(iy0,jy0);
 634             dz00             = _mm256_sub_ps(iz0,jz0);
 635
 636             /* Calculate squared distance and things based on it */
 637             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 638
 639             rinv00           = avx256_invsqrt_f(rsq00);
 640
 641             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 642
 643             /* Load parameters for j particles */
 644             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 645                                                                  charge+jnrC+0,charge+jnrD+0,
 646                                                                  charge+jnrE+0,charge+jnrF+0,
 647                                                                  charge+jnrG+0,charge+jnrH+0);
 648
 649             /**************************
 650              * CALCULATE INTERACTIONS *
 651              **************************/
 652
 653             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 654             {
 655
 656             r00              = _mm256_mul_ps(rsq00,rinv00);
 657
 658             /* Compute parameters for interactions between i and j atoms */
 659             qq00             = _mm256_mul_ps(iq0,jq0);
 660
 661             /* EWALD ELECTROSTATICS */
 662
 663             /* Analytical PME correction */
 664             zeta2            = _mm256_mul_ps(beta2,rsq00);
 665             rinv3            = _mm256_mul_ps(rinvsq00,rinv00);
 666             pmecorrF         = avx256_pmecorrF_f(zeta2);
 667             felec            = _mm256_add_ps( _mm256_mul_ps(pmecorrF,beta3), rinv3);
 668             felec            = _mm256_mul_ps(qq00,felec);
 669             pmecorrV         = avx256_pmecorrV_f(zeta2);
 670             pmecorrV         = _mm256_mul_ps(pmecorrV,beta);
 671             velec            = _mm256_sub_ps(rinv00,pmecorrV);
 672             velec            = _mm256_mul_ps(qq00,velec);
 673
 674             d                = _mm256_sub_ps(r00,rswitch);
 675             d                = _mm256_max_ps(d,_mm256_setzero_ps());
 676             d2               = _mm256_mul_ps(d,d);
 677             sw               = _mm256_add_ps(one,_mm256_mul_ps(d2,_mm256_mul_ps(d,_mm256_add_ps(swV3,_mm256_mul_ps(d,_mm256_add_ps(swV4,_mm256_mul_ps(d,swV5)))))));
 678
 679             dsw              = _mm256_mul_ps(d2,_mm256_add_ps(swF2,_mm256_mul_ps(d,_mm256_add_ps(swF3,_mm256_mul_ps(d,swF4)))));
 680
 681             /* Evaluate switch function */
 682             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 683             felec            = _mm256_sub_ps( _mm256_mul_ps(felec,sw) , _mm256_mul_ps(rinv00,_mm256_mul_ps(velec,dsw)) );
 684             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 685
 686             fscal            = felec;
 687
 688             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 689
 690             /* Calculate temporary vectorial force */
 691             tx               = _mm256_mul_ps(fscal,dx00);
 692             ty               = _mm256_mul_ps(fscal,dy00);
 693             tz               = _mm256_mul_ps(fscal,dz00);
 694
 695             /* Update vectorial force */
 696             fix0             = _mm256_add_ps(fix0,tx);
 697             fiy0             = _mm256_add_ps(fiy0,ty);
 698             fiz0             = _mm256_add_ps(fiz0,tz);
 699
 700             fjptrA             = f+j_coord_offsetA;
 701             fjptrB             = f+j_coord_offsetB;
 702             fjptrC             = f+j_coord_offsetC;
 703             fjptrD             = f+j_coord_offsetD;
 704             fjptrE             = f+j_coord_offsetE;
 705             fjptrF             = f+j_coord_offsetF;
 706             fjptrG             = f+j_coord_offsetG;
 707             fjptrH             = f+j_coord_offsetH;
 708             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 709
 710             }
 711
 712             /* Inner loop uses 105 flops */
 713         }
 714
 715         if(jidx<j_index_end)
 716         {
 717
 718             /* Get j neighbor index, and coordinate index */
 719             jnrlistA         = jjnr[jidx];
 720             jnrlistB         = jjnr[jidx+1];
 721             jnrlistC         = jjnr[jidx+2];
 722             jnrlistD         = jjnr[jidx+3];
 723             jnrlistE         = jjnr[jidx+4];
 724             jnrlistF         = jjnr[jidx+5];
 725             jnrlistG         = jjnr[jidx+6];
 726             jnrlistH         = jjnr[jidx+7];
 727             /* Sign of each element will be negative for non-real atoms.
 728              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 729              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 730              */
 731             dummy_mask = gmx_mm256_set_m128(gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx+4)),_mm_setzero_si128())),
 732                                             gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128())));
 733
 734             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 735             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 736             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 737             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 738             jnrE       = (jnrlistE>=0) ? jnrlistE : 0;
 739             jnrF       = (jnrlistF>=0) ? jnrlistF : 0;
 740             jnrG       = (jnrlistG>=0) ? jnrlistG : 0;
 741             jnrH       = (jnrlistH>=0) ? jnrlistH : 0;
 742             j_coord_offsetA  = DIM*jnrA;
 743             j_coord_offsetB  = DIM*jnrB;
 744             j_coord_offsetC  = DIM*jnrC;
 745             j_coord_offsetD  = DIM*jnrD;
 746             j_coord_offsetE  = DIM*jnrE;
 747             j_coord_offsetF  = DIM*jnrF;
 748             j_coord_offsetG  = DIM*jnrG;
 749             j_coord_offsetH  = DIM*jnrH;
 750
 751             /* load j atom coordinates */
 752             gmx_mm256_load_1rvec_8ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 753                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 754                                                  x+j_coord_offsetE,x+j_coord_offsetF,
 755                                                  x+j_coord_offsetG,x+j_coord_offsetH,
 756                                                  &jx0,&jy0,&jz0);
 757
 758             /* Calculate displacement vector */
 759             dx00             = _mm256_sub_ps(ix0,jx0);
 760             dy00             = _mm256_sub_ps(iy0,jy0);
 761             dz00             = _mm256_sub_ps(iz0,jz0);
 762
 763             /* Calculate squared distance and things based on it */
 764             rsq00            = gmx_mm256_calc_rsq_ps(dx00,dy00,dz00);
 765
 766             rinv00           = avx256_invsqrt_f(rsq00);
 767
 768             rinvsq00         = _mm256_mul_ps(rinv00,rinv00);
 769
 770             /* Load parameters for j particles */
 771             jq0              = gmx_mm256_load_8real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 772                                                                  charge+jnrC+0,charge+jnrD+0,
 773                                                                  charge+jnrE+0,charge+jnrF+0,
 774                                                                  charge+jnrG+0,charge+jnrH+0);
 775
 776             /**************************
 777              * CALCULATE INTERACTIONS *
 778              **************************/
 779
 780             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 781             {
 782
 783             r00              = _mm256_mul_ps(rsq00,rinv00);
 784             r00              = _mm256_andnot_ps(dummy_mask,r00);
 785
 786             /* Compute parameters for interactions between i and j atoms */
 787             qq00             = _mm256_mul_ps(iq0,jq0);
 788
 789             /* EWALD ELECTROSTATICS */
 790
 791             /* Analytical PME correction */
 792             zeta2            = _mm256_mul_ps(beta2,rsq00);
 793             rinv3            = _mm256_mul_ps(rinvsq00,rinv00);
 794             pmecorrF         = avx256_pmecorrF_f(zeta2);
 795             felec            = _mm256_add_ps( _mm256_mul_ps(pmecorrF,beta3), rinv3);
 796             felec            = _mm256_mul_ps(qq00,felec);
 797             pmecorrV         = avx256_pmecorrV_f(zeta2);
 798             pmecorrV         = _mm256_mul_ps(pmecorrV,beta);
 799             velec            = _mm256_sub_ps(rinv00,pmecorrV);
 800             velec            = _mm256_mul_ps(qq00,velec);
 801
 802             d                = _mm256_sub_ps(r00,rswitch);
 803             d                = _mm256_max_ps(d,_mm256_setzero_ps());
 804             d2               = _mm256_mul_ps(d,d);
 805             sw               = _mm256_add_ps(one,_mm256_mul_ps(d2,_mm256_mul_ps(d,_mm256_add_ps(swV3,_mm256_mul_ps(d,_mm256_add_ps(swV4,_mm256_mul_ps(d,swV5)))))));
 806
 807             dsw              = _mm256_mul_ps(d2,_mm256_add_ps(swF2,_mm256_mul_ps(d,_mm256_add_ps(swF3,_mm256_mul_ps(d,swF4)))));
 808
 809             /* Evaluate switch function */
 810             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 811             felec            = _mm256_sub_ps( _mm256_mul_ps(felec,sw) , _mm256_mul_ps(rinv00,_mm256_mul_ps(velec,dsw)) );
 812             cutoff_mask      = _mm256_cmp_ps(rsq00,rcutoff2,_CMP_LT_OQ);
 813
 814             fscal            = felec;
 815
 816             fscal            = _mm256_and_ps(fscal,cutoff_mask);
 817
 818             fscal            = _mm256_andnot_ps(dummy_mask,fscal);
 819
 820             /* Calculate temporary vectorial force */
 821             tx               = _mm256_mul_ps(fscal,dx00);
 822             ty               = _mm256_mul_ps(fscal,dy00);
 823             tz               = _mm256_mul_ps(fscal,dz00);
 824
 825             /* Update vectorial force */
 826             fix0             = _mm256_add_ps(fix0,tx);
 827             fiy0             = _mm256_add_ps(fiy0,ty);
 828             fiz0             = _mm256_add_ps(fiz0,tz);
 829
 830             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 831             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 832             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 833             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 834             fjptrE             = (jnrlistE>=0) ? f+j_coord_offsetE : scratch;
 835             fjptrF             = (jnrlistF>=0) ? f+j_coord_offsetF : scratch;
 836             fjptrG             = (jnrlistG>=0) ? f+j_coord_offsetG : scratch;
 837             fjptrH             = (jnrlistH>=0) ? f+j_coord_offsetH : scratch;
 838             gmx_mm256_decrement_1rvec_8ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,fjptrE,fjptrF,fjptrG,fjptrH,tx,ty,tz);
 839
 840             }
 841
 842             /* Inner loop uses 106 flops */
 843         }
 844
 845         /* End of innermost loop */
 846
 847         gmx_mm256_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 848                                                  f+i_coord_offset,fshift+i_shift_offset);
 849
 850         /* Increment number of inner iterations */
 851         inneriter                  += j_index_end - j_index_start;
 852
 853         /* Outer loop uses 7 flops */
 854     }
 855
 856     /* Increment number of outer iterations */
 857     outeriter        += nri;
 858
 859     /* Update outer/inner flops */
 860
 861     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*106);
 862 }