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) 2012,2013,2014, 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 "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/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_ElecEwSw_VdwNone_GeomP1P1_VF_avx_256_double
  52  * Electrostatics interaction: Ewald
  53  * VdW interaction:            None
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecEwSw_VdwNone_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     __m128i          ewitab;
  92     __m256d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  93     __m256d          beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
  94     real             *ewtab;
  95     __m256d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  96     real             rswitch_scalar,d_scalar;
  97     __m256d          dummy_mask,cutoff_mask;
  98     __m128           tmpmask0,tmpmask1;
  99     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 100     __m256d          one     = _mm256_set1_pd(1.0);
 101     __m256d          two     = _mm256_set1_pd(2.0);
 102     x                = xx[0];
 103     f                = ff[0];
 104
 105     nri              = nlist->nri;
 106     iinr             = nlist->iinr;
 107     jindex           = nlist->jindex;
 108     jjnr             = nlist->jjnr;
 109     shiftidx         = nlist->shift;
 110     gid              = nlist->gid;
 111     shiftvec         = fr->shift_vec[0];
 112     fshift           = fr->fshift[0];
 113     facel            = _mm256_set1_pd(fr->epsfac);
 114     charge           = mdatoms->chargeA;
 115
 116     sh_ewald         = _mm256_set1_pd(fr->ic->sh_ewald);
 117     beta             = _mm256_set1_pd(fr->ic->ewaldcoeff_q);
 118     beta2            = _mm256_mul_pd(beta,beta);
 119     beta3            = _mm256_mul_pd(beta,beta2);
 120
 121     ewtab            = fr->ic->tabq_coul_FDV0;
 122     ewtabscale       = _mm256_set1_pd(fr->ic->tabq_scale);
 123     ewtabhalfspace   = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
 124
 125     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 126     rcutoff_scalar   = fr->rcoulomb;
 127     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 128     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 129
 130     rswitch_scalar   = fr->rcoulomb_switch;
 131     rswitch          = _mm256_set1_pd(rswitch_scalar);
 132     /* Setup switch parameters */
 133     d_scalar         = rcutoff_scalar-rswitch_scalar;
 134     d                = _mm256_set1_pd(d_scalar);
 135     swV3             = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 136     swV4             = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 137     swV5             = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 138     swF2             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 139     swF3             = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 140     swF4             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 141
 142     /* Avoid stupid compiler warnings */
 143     jnrA = jnrB = jnrC = jnrD = 0;
 144     j_coord_offsetA = 0;
 145     j_coord_offsetB = 0;
 146     j_coord_offsetC = 0;
 147     j_coord_offsetD = 0;
 148
 149     outeriter        = 0;
 150     inneriter        = 0;
 151
 152     for(iidx=0;iidx<4*DIM;iidx++)
 153     {
 154         scratch[iidx] = 0.0;
 155     }
 156
 157     /* Start outer loop over neighborlists */
 158     for(iidx=0; iidx<nri; iidx++)
 159     {
 160         /* Load shift vector for this list */
 161         i_shift_offset   = DIM*shiftidx[iidx];
 162
 163         /* Load limits for loop over neighbors */
 164         j_index_start    = jindex[iidx];
 165         j_index_end      = jindex[iidx+1];
 166
 167         /* Get outer coordinate index */
 168         inr              = iinr[iidx];
 169         i_coord_offset   = DIM*inr;
 170
 171         /* Load i particle coords and add shift vector */
 172         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 173
 174         fix0             = _mm256_setzero_pd();
 175         fiy0             = _mm256_setzero_pd();
 176         fiz0             = _mm256_setzero_pd();
 177
 178         /* Load parameters for i particles */
 179         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 180
 181         /* Reset potential sums */
 182         velecsum         = _mm256_setzero_pd();
 183
 184         /* Start inner kernel loop */
 185         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 186         {
 187
 188             /* Get j neighbor index, and coordinate index */
 189             jnrA             = jjnr[jidx];
 190             jnrB             = jjnr[jidx+1];
 191             jnrC             = jjnr[jidx+2];
 192             jnrD             = jjnr[jidx+3];
 193             j_coord_offsetA  = DIM*jnrA;
 194             j_coord_offsetB  = DIM*jnrB;
 195             j_coord_offsetC  = DIM*jnrC;
 196             j_coord_offsetD  = DIM*jnrD;
 197
 198             /* load j atom coordinates */
 199             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 200                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 201                                                  &jx0,&jy0,&jz0);
 202
 203             /* Calculate displacement vector */
 204             dx00             = _mm256_sub_pd(ix0,jx0);
 205             dy00             = _mm256_sub_pd(iy0,jy0);
 206             dz00             = _mm256_sub_pd(iz0,jz0);
 207
 208             /* Calculate squared distance and things based on it */
 209             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 210
 211             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 212
 213             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 214
 215             /* Load parameters for j particles */
 216             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 217                                                                  charge+jnrC+0,charge+jnrD+0);
 218
 219             /**************************
 220              * CALCULATE INTERACTIONS *
 221              **************************/
 222
 223             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 224             {
 225
 226             r00              = _mm256_mul_pd(rsq00,rinv00);
 227
 228             /* Compute parameters for interactions between i and j atoms */
 229             qq00             = _mm256_mul_pd(iq0,jq0);
 230
 231             /* EWALD ELECTROSTATICS */
 232
 233             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 234             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 235             ewitab           = _mm256_cvttpd_epi32(ewrt);
 236             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 237             ewitab           = _mm_slli_epi32(ewitab,2);
 238             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 239             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 240             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 241             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 242             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 243             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 244             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 245             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
 246             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 247
 248             d                = _mm256_sub_pd(r00,rswitch);
 249             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 250             d2               = _mm256_mul_pd(d,d);
 251             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)))))));
 252
 253             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 254
 255             /* Evaluate switch function */
 256             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 257             felec            = _mm256_sub_pd( _mm256_mul_pd(felec,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(velec,dsw)) );
 258             velec            = _mm256_mul_pd(velec,sw);
 259             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 260
 261             /* Update potential sum for this i atom from the interaction with this j atom. */
 262             velec            = _mm256_and_pd(velec,cutoff_mask);
 263             velecsum         = _mm256_add_pd(velecsum,velec);
 264
 265             fscal            = felec;
 266
 267             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 268
 269             /* Calculate temporary vectorial force */
 270             tx               = _mm256_mul_pd(fscal,dx00);
 271             ty               = _mm256_mul_pd(fscal,dy00);
 272             tz               = _mm256_mul_pd(fscal,dz00);
 273
 274             /* Update vectorial force */
 275             fix0             = _mm256_add_pd(fix0,tx);
 276             fiy0             = _mm256_add_pd(fiy0,ty);
 277             fiz0             = _mm256_add_pd(fiz0,tz);
 278
 279             fjptrA             = f+j_coord_offsetA;
 280             fjptrB             = f+j_coord_offsetB;
 281             fjptrC             = f+j_coord_offsetC;
 282             fjptrD             = f+j_coord_offsetD;
 283             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 284
 285             }
 286
 287             /* Inner loop uses 65 flops */
 288         }
 289
 290         if(jidx<j_index_end)
 291         {
 292
 293             /* Get j neighbor index, and coordinate index */
 294             jnrlistA         = jjnr[jidx];
 295             jnrlistB         = jjnr[jidx+1];
 296             jnrlistC         = jjnr[jidx+2];
 297             jnrlistD         = jjnr[jidx+3];
 298             /* Sign of each element will be negative for non-real atoms.
 299              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 300              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 301              */
 302             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 303
 304             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 305             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 306             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 307
 308             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 309             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 310             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 311             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 312             j_coord_offsetA  = DIM*jnrA;
 313             j_coord_offsetB  = DIM*jnrB;
 314             j_coord_offsetC  = DIM*jnrC;
 315             j_coord_offsetD  = DIM*jnrD;
 316
 317             /* load j atom coordinates */
 318             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 319                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 320                                                  &jx0,&jy0,&jz0);
 321
 322             /* Calculate displacement vector */
 323             dx00             = _mm256_sub_pd(ix0,jx0);
 324             dy00             = _mm256_sub_pd(iy0,jy0);
 325             dz00             = _mm256_sub_pd(iz0,jz0);
 326
 327             /* Calculate squared distance and things based on it */
 328             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 329
 330             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 331
 332             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 333
 334             /* Load parameters for j particles */
 335             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 336                                                                  charge+jnrC+0,charge+jnrD+0);
 337
 338             /**************************
 339              * CALCULATE INTERACTIONS *
 340              **************************/
 341
 342             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 343             {
 344
 345             r00              = _mm256_mul_pd(rsq00,rinv00);
 346             r00              = _mm256_andnot_pd(dummy_mask,r00);
 347
 348             /* Compute parameters for interactions between i and j atoms */
 349             qq00             = _mm256_mul_pd(iq0,jq0);
 350
 351             /* EWALD ELECTROSTATICS */
 352
 353             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 354             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 355             ewitab           = _mm256_cvttpd_epi32(ewrt);
 356             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 357             ewitab           = _mm_slli_epi32(ewitab,2);
 358             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 359             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 360             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 361             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 362             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 363             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 364             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 365             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
 366             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 367
 368             d                = _mm256_sub_pd(r00,rswitch);
 369             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 370             d2               = _mm256_mul_pd(d,d);
 371             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)))))));
 372
 373             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 374
 375             /* Evaluate switch function */
 376             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 377             felec            = _mm256_sub_pd( _mm256_mul_pd(felec,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(velec,dsw)) );
 378             velec            = _mm256_mul_pd(velec,sw);
 379             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 380
 381             /* Update potential sum for this i atom from the interaction with this j atom. */
 382             velec            = _mm256_and_pd(velec,cutoff_mask);
 383             velec            = _mm256_andnot_pd(dummy_mask,velec);
 384             velecsum         = _mm256_add_pd(velecsum,velec);
 385
 386             fscal            = felec;
 387
 388             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 389
 390             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 391
 392             /* Calculate temporary vectorial force */
 393             tx               = _mm256_mul_pd(fscal,dx00);
 394             ty               = _mm256_mul_pd(fscal,dy00);
 395             tz               = _mm256_mul_pd(fscal,dz00);
 396
 397             /* Update vectorial force */
 398             fix0             = _mm256_add_pd(fix0,tx);
 399             fiy0             = _mm256_add_pd(fiy0,ty);
 400             fiz0             = _mm256_add_pd(fiz0,tz);
 401
 402             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 403             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 404             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 405             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 406             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 407
 408             }
 409
 410             /* Inner loop uses 66 flops */
 411         }
 412
 413         /* End of innermost loop */
 414
 415         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 416                                                  f+i_coord_offset,fshift+i_shift_offset);
 417
 418         ggid                        = gid[iidx];
 419         /* Update potential energies */
 420         gmx_mm256_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 421
 422         /* Increment number of inner iterations */
 423         inneriter                  += j_index_end - j_index_start;
 424
 425         /* Outer loop uses 8 flops */
 426     }
 427
 428     /* Increment number of outer iterations */
 429     outeriter        += nri;
 430
 431     /* Update outer/inner flops */
 432
 433     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VF,outeriter*8 + inneriter*66);
 434 }
 435 /*
 436  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwNone_GeomP1P1_F_avx_256_double
 437  * Electrostatics interaction: Ewald
 438  * VdW interaction:            None
 439  * Geometry:                   Particle-Particle
 440  * Calculate force/pot:        Force
 441  */
 442 void
 443 nb_kernel_ElecEwSw_VdwNone_GeomP1P1_F_avx_256_double
 444                     (t_nblist                    * gmx_restrict       nlist,
 445                      rvec                        * gmx_restrict          xx,
 446                      rvec                        * gmx_restrict          ff,
 447                      t_forcerec                  * gmx_restrict          fr,
 448                      t_mdatoms                   * gmx_restrict     mdatoms,
 449                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 450                      t_nrnb                      * gmx_restrict        nrnb)
 451 {
 452     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 453      * just 0 for non-waters.
 454      * Suffixes A,B,C,D refer to j loop unrolling done with AVX, e.g. for the four different
 455      * jnr indices corresponding to data put in the four positions in the SIMD register.
 456      */
 457     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 458     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 459     int              jnrA,jnrB,jnrC,jnrD;
 460     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 461     int              jnrlistE,jnrlistF,jnrlistG,jnrlistH;
 462     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 463     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 464     real             rcutoff_scalar;
 465     real             *shiftvec,*fshift,*x,*f;
 466     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 467     real             scratch[4*DIM];
 468     __m256d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 469     real *           vdwioffsetptr0;
 470     __m256d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 471     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 472     __m256d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 473     __m256d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 474     __m256d          velec,felec,velecsum,facel,crf,krf,krf2;
 475     real             *charge;
 476     __m128i          ewitab;
 477     __m256d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 478     __m256d          beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 479     real             *ewtab;
 480     __m256d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 481     real             rswitch_scalar,d_scalar;
 482     __m256d          dummy_mask,cutoff_mask;
 483     __m128           tmpmask0,tmpmask1;
 484     __m256d          signbit = _mm256_castsi256_pd( _mm256_set1_epi32(0x80000000) );
 485     __m256d          one     = _mm256_set1_pd(1.0);
 486     __m256d          two     = _mm256_set1_pd(2.0);
 487     x                = xx[0];
 488     f                = ff[0];
 489
 490     nri              = nlist->nri;
 491     iinr             = nlist->iinr;
 492     jindex           = nlist->jindex;
 493     jjnr             = nlist->jjnr;
 494     shiftidx         = nlist->shift;
 495     gid              = nlist->gid;
 496     shiftvec         = fr->shift_vec[0];
 497     fshift           = fr->fshift[0];
 498     facel            = _mm256_set1_pd(fr->epsfac);
 499     charge           = mdatoms->chargeA;
 500
 501     sh_ewald         = _mm256_set1_pd(fr->ic->sh_ewald);
 502     beta             = _mm256_set1_pd(fr->ic->ewaldcoeff_q);
 503     beta2            = _mm256_mul_pd(beta,beta);
 504     beta3            = _mm256_mul_pd(beta,beta2);
 505
 506     ewtab            = fr->ic->tabq_coul_FDV0;
 507     ewtabscale       = _mm256_set1_pd(fr->ic->tabq_scale);
 508     ewtabhalfspace   = _mm256_set1_pd(0.5/fr->ic->tabq_scale);
 509
 510     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 511     rcutoff_scalar   = fr->rcoulomb;
 512     rcutoff          = _mm256_set1_pd(rcutoff_scalar);
 513     rcutoff2         = _mm256_mul_pd(rcutoff,rcutoff);
 514
 515     rswitch_scalar   = fr->rcoulomb_switch;
 516     rswitch          = _mm256_set1_pd(rswitch_scalar);
 517     /* Setup switch parameters */
 518     d_scalar         = rcutoff_scalar-rswitch_scalar;
 519     d                = _mm256_set1_pd(d_scalar);
 520     swV3             = _mm256_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 521     swV4             = _mm256_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 522     swV5             = _mm256_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 523     swF2             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 524     swF3             = _mm256_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 525     swF4             = _mm256_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 526
 527     /* Avoid stupid compiler warnings */
 528     jnrA = jnrB = jnrC = jnrD = 0;
 529     j_coord_offsetA = 0;
 530     j_coord_offsetB = 0;
 531     j_coord_offsetC = 0;
 532     j_coord_offsetD = 0;
 533
 534     outeriter        = 0;
 535     inneriter        = 0;
 536
 537     for(iidx=0;iidx<4*DIM;iidx++)
 538     {
 539         scratch[iidx] = 0.0;
 540     }
 541
 542     /* Start outer loop over neighborlists */
 543     for(iidx=0; iidx<nri; iidx++)
 544     {
 545         /* Load shift vector for this list */
 546         i_shift_offset   = DIM*shiftidx[iidx];
 547
 548         /* Load limits for loop over neighbors */
 549         j_index_start    = jindex[iidx];
 550         j_index_end      = jindex[iidx+1];
 551
 552         /* Get outer coordinate index */
 553         inr              = iinr[iidx];
 554         i_coord_offset   = DIM*inr;
 555
 556         /* Load i particle coords and add shift vector */
 557         gmx_mm256_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 558
 559         fix0             = _mm256_setzero_pd();
 560         fiy0             = _mm256_setzero_pd();
 561         fiz0             = _mm256_setzero_pd();
 562
 563         /* Load parameters for i particles */
 564         iq0              = _mm256_mul_pd(facel,_mm256_set1_pd(charge[inr+0]));
 565
 566         /* Start inner kernel loop */
 567         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 568         {
 569
 570             /* Get j neighbor index, and coordinate index */
 571             jnrA             = jjnr[jidx];
 572             jnrB             = jjnr[jidx+1];
 573             jnrC             = jjnr[jidx+2];
 574             jnrD             = jjnr[jidx+3];
 575             j_coord_offsetA  = DIM*jnrA;
 576             j_coord_offsetB  = DIM*jnrB;
 577             j_coord_offsetC  = DIM*jnrC;
 578             j_coord_offsetD  = DIM*jnrD;
 579
 580             /* load j atom coordinates */
 581             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 582                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 583                                                  &jx0,&jy0,&jz0);
 584
 585             /* Calculate displacement vector */
 586             dx00             = _mm256_sub_pd(ix0,jx0);
 587             dy00             = _mm256_sub_pd(iy0,jy0);
 588             dz00             = _mm256_sub_pd(iz0,jz0);
 589
 590             /* Calculate squared distance and things based on it */
 591             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 592
 593             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 594
 595             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 596
 597             /* Load parameters for j particles */
 598             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 599                                                                  charge+jnrC+0,charge+jnrD+0);
 600
 601             /**************************
 602              * CALCULATE INTERACTIONS *
 603              **************************/
 604
 605             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 606             {
 607
 608             r00              = _mm256_mul_pd(rsq00,rinv00);
 609
 610             /* Compute parameters for interactions between i and j atoms */
 611             qq00             = _mm256_mul_pd(iq0,jq0);
 612
 613             /* EWALD ELECTROSTATICS */
 614
 615             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 616             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 617             ewitab           = _mm256_cvttpd_epi32(ewrt);
 618             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 619             ewitab           = _mm_slli_epi32(ewitab,2);
 620             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 621             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 622             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 623             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 624             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 625             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 626             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 627             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
 628             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 629
 630             d                = _mm256_sub_pd(r00,rswitch);
 631             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 632             d2               = _mm256_mul_pd(d,d);
 633             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)))))));
 634
 635             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 636
 637             /* Evaluate switch function */
 638             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 639             felec            = _mm256_sub_pd( _mm256_mul_pd(felec,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(velec,dsw)) );
 640             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 641
 642             fscal            = felec;
 643
 644             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 645
 646             /* Calculate temporary vectorial force */
 647             tx               = _mm256_mul_pd(fscal,dx00);
 648             ty               = _mm256_mul_pd(fscal,dy00);
 649             tz               = _mm256_mul_pd(fscal,dz00);
 650
 651             /* Update vectorial force */
 652             fix0             = _mm256_add_pd(fix0,tx);
 653             fiy0             = _mm256_add_pd(fiy0,ty);
 654             fiz0             = _mm256_add_pd(fiz0,tz);
 655
 656             fjptrA             = f+j_coord_offsetA;
 657             fjptrB             = f+j_coord_offsetB;
 658             fjptrC             = f+j_coord_offsetC;
 659             fjptrD             = f+j_coord_offsetD;
 660             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 661
 662             }
 663
 664             /* Inner loop uses 62 flops */
 665         }
 666
 667         if(jidx<j_index_end)
 668         {
 669
 670             /* Get j neighbor index, and coordinate index */
 671             jnrlistA         = jjnr[jidx];
 672             jnrlistB         = jjnr[jidx+1];
 673             jnrlistC         = jjnr[jidx+2];
 674             jnrlistD         = jjnr[jidx+3];
 675             /* Sign of each element will be negative for non-real atoms.
 676              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 677              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
 678              */
 679             tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 680
 681             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
 682             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
 683             dummy_mask = _mm256_castps_pd(gmx_mm256_set_m128(tmpmask1,tmpmask0));
 684
 685             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 686             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 687             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 688             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 689             j_coord_offsetA  = DIM*jnrA;
 690             j_coord_offsetB  = DIM*jnrB;
 691             j_coord_offsetC  = DIM*jnrC;
 692             j_coord_offsetD  = DIM*jnrD;
 693
 694             /* load j atom coordinates */
 695             gmx_mm256_load_1rvec_4ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 696                                                  x+j_coord_offsetC,x+j_coord_offsetD,
 697                                                  &jx0,&jy0,&jz0);
 698
 699             /* Calculate displacement vector */
 700             dx00             = _mm256_sub_pd(ix0,jx0);
 701             dy00             = _mm256_sub_pd(iy0,jy0);
 702             dz00             = _mm256_sub_pd(iz0,jz0);
 703
 704             /* Calculate squared distance and things based on it */
 705             rsq00            = gmx_mm256_calc_rsq_pd(dx00,dy00,dz00);
 706
 707             rinv00           = gmx_mm256_invsqrt_pd(rsq00);
 708
 709             rinvsq00         = _mm256_mul_pd(rinv00,rinv00);
 710
 711             /* Load parameters for j particles */
 712             jq0              = gmx_mm256_load_4real_swizzle_pd(charge+jnrA+0,charge+jnrB+0,
 713                                                                  charge+jnrC+0,charge+jnrD+0);
 714
 715             /**************************
 716              * CALCULATE INTERACTIONS *
 717              **************************/
 718
 719             if (gmx_mm256_any_lt(rsq00,rcutoff2))
 720             {
 721
 722             r00              = _mm256_mul_pd(rsq00,rinv00);
 723             r00              = _mm256_andnot_pd(dummy_mask,r00);
 724
 725             /* Compute parameters for interactions between i and j atoms */
 726             qq00             = _mm256_mul_pd(iq0,jq0);
 727
 728             /* EWALD ELECTROSTATICS */
 729
 730             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 731             ewrt             = _mm256_mul_pd(r00,ewtabscale);
 732             ewitab           = _mm256_cvttpd_epi32(ewrt);
 733             eweps            = _mm256_sub_pd(ewrt,_mm256_round_pd(ewrt, _MM_FROUND_FLOOR));
 734             ewitab           = _mm_slli_epi32(ewitab,2);
 735             ewtabF           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 736             ewtabD           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 737             ewtabV           = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,2) );
 738             ewtabFn          = _mm256_load_pd( ewtab + _mm_extract_epi32(ewitab,3) );
 739             GMX_MM256_FULLTRANSPOSE4_PD(ewtabF,ewtabD,ewtabV,ewtabFn);
 740             felec            = _mm256_add_pd(ewtabF,_mm256_mul_pd(eweps,ewtabD));
 741             velec            = _mm256_sub_pd(ewtabV,_mm256_mul_pd(_mm256_mul_pd(ewtabhalfspace,eweps),_mm256_add_pd(ewtabF,felec)));
 742             velec            = _mm256_mul_pd(qq00,_mm256_sub_pd(rinv00,velec));
 743             felec            = _mm256_mul_pd(_mm256_mul_pd(qq00,rinv00),_mm256_sub_pd(rinvsq00,felec));
 744
 745             d                = _mm256_sub_pd(r00,rswitch);
 746             d                = _mm256_max_pd(d,_mm256_setzero_pd());
 747             d2               = _mm256_mul_pd(d,d);
 748             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)))))));
 749
 750             dsw              = _mm256_mul_pd(d2,_mm256_add_pd(swF2,_mm256_mul_pd(d,_mm256_add_pd(swF3,_mm256_mul_pd(d,swF4)))));
 751
 752             /* Evaluate switch function */
 753             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 754             felec            = _mm256_sub_pd( _mm256_mul_pd(felec,sw) , _mm256_mul_pd(rinv00,_mm256_mul_pd(velec,dsw)) );
 755             cutoff_mask      = _mm256_cmp_pd(rsq00,rcutoff2,_CMP_LT_OQ);
 756
 757             fscal            = felec;
 758
 759             fscal            = _mm256_and_pd(fscal,cutoff_mask);
 760
 761             fscal            = _mm256_andnot_pd(dummy_mask,fscal);
 762
 763             /* Calculate temporary vectorial force */
 764             tx               = _mm256_mul_pd(fscal,dx00);
 765             ty               = _mm256_mul_pd(fscal,dy00);
 766             tz               = _mm256_mul_pd(fscal,dz00);
 767
 768             /* Update vectorial force */
 769             fix0             = _mm256_add_pd(fix0,tx);
 770             fiy0             = _mm256_add_pd(fiy0,ty);
 771             fiz0             = _mm256_add_pd(fiz0,tz);
 772
 773             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 774             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 775             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 776             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 777             gmx_mm256_decrement_1rvec_4ptr_swizzle_pd(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 778
 779             }
 780
 781             /* Inner loop uses 63 flops */
 782         }
 783
 784         /* End of innermost loop */
 785
 786         gmx_mm256_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 787                                                  f+i_coord_offset,fshift+i_shift_offset);
 788
 789         /* Increment number of inner iterations */
 790         inneriter                  += j_index_end - j_index_start;
 791
 792         /* Outer loop uses 7 flops */
 793     }
 794
 795     /* Increment number of outer iterations */
 796     outeriter        += nri;
 797
 798     /* Update outer/inner flops */
 799
 800     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_F,outeriter*7 + inneriter*63);
 801 }