src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecEwSw_VdwLJSw_GeomP1P1_avx_128_fma_double.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_128_fma_double 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_128_fma_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwLJSw_GeomP1P1_VF_avx_128_fma_double
  51  * Electrostatics interaction: Ewald
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecEwSw_VdwLJSw_GeomP1P1_VF_avx_128_fma_double
  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 refer to j loop unrolling done with SSE double precision, e.g. for the two 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;
  74     int              j_coord_offsetA,j_coord_offsetB;
  75     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  76     real             rcutoff_scalar;
  77     real             *shiftvec,*fshift,*x,*f;
  78     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  79     int              vdwioffset0;
  80     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  81     int              vdwjidx0A,vdwjidx0B;
  82     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  83     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  84     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  85     real             *charge;
  86     int              nvdwtype;
  87     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  88     int              *vdwtype;
  89     real             *vdwparam;
  90     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
  91     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
  92     __m128i          ewitab;
  93     __m128d          ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  94     real             *ewtab;
  95     __m128d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  96     real             rswitch_scalar,d_scalar;
  97     __m128d          dummy_mask,cutoff_mask;
  98     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
  99     __m128d          one     = _mm_set1_pd(1.0);
 100     __m128d          two     = _mm_set1_pd(2.0);
 101     x                = xx[0];
 102     f                = ff[0];
 103
 104     nri              = nlist->nri;
 105     iinr             = nlist->iinr;
 106     jindex           = nlist->jindex;
 107     jjnr             = nlist->jjnr;
 108     shiftidx         = nlist->shift;
 109     gid              = nlist->gid;
 110     shiftvec         = fr->shift_vec[0];
 111     fshift           = fr->fshift[0];
 112     facel            = _mm_set1_pd(fr->ic->epsfac);
 113     charge           = mdatoms->chargeA;
 114     nvdwtype         = fr->ntype;
 115     vdwparam         = fr->nbfp;
 116     vdwtype          = mdatoms->typeA;
 117
 118     sh_ewald         = _mm_set1_pd(fr->ic->sh_ewald);
 119     ewtab            = fr->ic->tabq_coul_FDV0;
 120     ewtabscale       = _mm_set1_pd(fr->ic->tabq_scale);
 121     ewtabhalfspace   = _mm_set1_pd(0.5/fr->ic->tabq_scale);
 122
 123     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 124     rcutoff_scalar   = fr->ic->rcoulomb;
 125     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 126     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 127
 128     rswitch_scalar   = fr->ic->rcoulomb_switch;
 129     rswitch          = _mm_set1_pd(rswitch_scalar);
 130     /* Setup switch parameters */
 131     d_scalar         = rcutoff_scalar-rswitch_scalar;
 132     d                = _mm_set1_pd(d_scalar);
 133     swV3             = _mm_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 134     swV4             = _mm_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 135     swV5             = _mm_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 136     swF2             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 137     swF3             = _mm_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 138     swF4             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 139
 140     /* Avoid stupid compiler warnings */
 141     jnrA = jnrB = 0;
 142     j_coord_offsetA = 0;
 143     j_coord_offsetB = 0;
 144
 145     outeriter        = 0;
 146     inneriter        = 0;
 147
 148     /* Start outer loop over neighborlists */
 149     for(iidx=0; iidx<nri; iidx++)
 150     {
 151         /* Load shift vector for this list */
 152         i_shift_offset   = DIM*shiftidx[iidx];
 153
 154         /* Load limits for loop over neighbors */
 155         j_index_start    = jindex[iidx];
 156         j_index_end      = jindex[iidx+1];
 157
 158         /* Get outer coordinate index */
 159         inr              = iinr[iidx];
 160         i_coord_offset   = DIM*inr;
 161
 162         /* Load i particle coords and add shift vector */
 163         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 164
 165         fix0             = _mm_setzero_pd();
 166         fiy0             = _mm_setzero_pd();
 167         fiz0             = _mm_setzero_pd();
 168
 169         /* Load parameters for i particles */
 170         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 171         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 172
 173         /* Reset potential sums */
 174         velecsum         = _mm_setzero_pd();
 175         vvdwsum          = _mm_setzero_pd();
 176
 177         /* Start inner kernel loop */
 178         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 179         {
 180
 181             /* Get j neighbor index, and coordinate index */
 182             jnrA             = jjnr[jidx];
 183             jnrB             = jjnr[jidx+1];
 184             j_coord_offsetA  = DIM*jnrA;
 185             j_coord_offsetB  = DIM*jnrB;
 186
 187             /* load j atom coordinates */
 188             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 189                                               &jx0,&jy0,&jz0);
 190
 191             /* Calculate displacement vector */
 192             dx00             = _mm_sub_pd(ix0,jx0);
 193             dy00             = _mm_sub_pd(iy0,jy0);
 194             dz00             = _mm_sub_pd(iz0,jz0);
 195
 196             /* Calculate squared distance and things based on it */
 197             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 198
 199             rinv00           = avx128fma_invsqrt_d(rsq00);
 200
 201             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 202
 203             /* Load parameters for j particles */
 204             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 205             vdwjidx0A        = 2*vdwtype[jnrA+0];
 206             vdwjidx0B        = 2*vdwtype[jnrB+0];
 207
 208             /**************************
 209              * CALCULATE INTERACTIONS *
 210              **************************/
 211
 212             if (gmx_mm_any_lt(rsq00,rcutoff2))
 213             {
 214
 215             r00              = _mm_mul_pd(rsq00,rinv00);
 216
 217             /* Compute parameters for interactions between i and j atoms */
 218             qq00             = _mm_mul_pd(iq0,jq0);
 219             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 220                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 221
 222             /* EWALD ELECTROSTATICS */
 223
 224             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 225             ewrt             = _mm_mul_pd(r00,ewtabscale);
 226             ewitab           = _mm_cvttpd_epi32(ewrt);
 227 #ifdef __XOP__
 228             eweps            = _mm_frcz_pd(ewrt);
 229 #else
 230             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 231 #endif
 232             twoeweps         = _mm_add_pd(eweps,eweps);
 233             ewitab           = _mm_slli_epi32(ewitab,2);
 234             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 235             ewtabD           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 236             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 237             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 238             ewtabFn          = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,1) +2);
 239             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 240             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 241             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 242             velec            = _mm_mul_pd(qq00,_mm_sub_pd(rinv00,velec));
 243             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 244
 245             /* LENNARD-JONES DISPERSION/REPULSION */
 246
 247             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 248             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 249             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 250             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 251             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 252
 253             d                = _mm_sub_pd(r00,rswitch);
 254             d                = _mm_max_pd(d,_mm_setzero_pd());
 255             d2               = _mm_mul_pd(d,d);
 256             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 257
 258             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 259
 260             /* Evaluate switch function */
 261             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 262             felec            = _mm_msub_pd( felec,sw , _mm_mul_pd(rinv00,_mm_mul_pd(velec,dsw)) );
 263             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 264             velec            = _mm_mul_pd(velec,sw);
 265             vvdw             = _mm_mul_pd(vvdw,sw);
 266             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 267
 268             /* Update potential sum for this i atom from the interaction with this j atom. */
 269             velec            = _mm_and_pd(velec,cutoff_mask);
 270             velecsum         = _mm_add_pd(velecsum,velec);
 271             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 272             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 273
 274             fscal            = _mm_add_pd(felec,fvdw);
 275
 276             fscal            = _mm_and_pd(fscal,cutoff_mask);
 277
 278             /* Update vectorial force */
 279             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 280             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 281             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 282
 283             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 284                                                    _mm_mul_pd(dx00,fscal),
 285                                                    _mm_mul_pd(dy00,fscal),
 286                                                    _mm_mul_pd(dz00,fscal));
 287
 288             }
 289
 290             /* Inner loop uses 86 flops */
 291         }
 292
 293         if(jidx<j_index_end)
 294         {
 295
 296             jnrA             = jjnr[jidx];
 297             j_coord_offsetA  = DIM*jnrA;
 298
 299             /* load j atom coordinates */
 300             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 301                                               &jx0,&jy0,&jz0);
 302
 303             /* Calculate displacement vector */
 304             dx00             = _mm_sub_pd(ix0,jx0);
 305             dy00             = _mm_sub_pd(iy0,jy0);
 306             dz00             = _mm_sub_pd(iz0,jz0);
 307
 308             /* Calculate squared distance and things based on it */
 309             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 310
 311             rinv00           = avx128fma_invsqrt_d(rsq00);
 312
 313             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 314
 315             /* Load parameters for j particles */
 316             jq0              = _mm_load_sd(charge+jnrA+0);
 317             vdwjidx0A        = 2*vdwtype[jnrA+0];
 318
 319             /**************************
 320              * CALCULATE INTERACTIONS *
 321              **************************/
 322
 323             if (gmx_mm_any_lt(rsq00,rcutoff2))
 324             {
 325
 326             r00              = _mm_mul_pd(rsq00,rinv00);
 327
 328             /* Compute parameters for interactions between i and j atoms */
 329             qq00             = _mm_mul_pd(iq0,jq0);
 330             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 331
 332             /* EWALD ELECTROSTATICS */
 333
 334             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 335             ewrt             = _mm_mul_pd(r00,ewtabscale);
 336             ewitab           = _mm_cvttpd_epi32(ewrt);
 337 #ifdef __XOP__
 338             eweps            = _mm_frcz_pd(ewrt);
 339 #else
 340             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 341 #endif
 342             twoeweps         = _mm_add_pd(eweps,eweps);
 343             ewitab           = _mm_slli_epi32(ewitab,2);
 344             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 345             ewtabD           = _mm_setzero_pd();
 346             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 347             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 348             ewtabFn          = _mm_setzero_pd();
 349             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 350             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 351             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 352             velec            = _mm_mul_pd(qq00,_mm_sub_pd(rinv00,velec));
 353             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 354
 355             /* LENNARD-JONES DISPERSION/REPULSION */
 356
 357             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 358             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 359             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 360             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 361             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 362
 363             d                = _mm_sub_pd(r00,rswitch);
 364             d                = _mm_max_pd(d,_mm_setzero_pd());
 365             d2               = _mm_mul_pd(d,d);
 366             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 367
 368             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 369
 370             /* Evaluate switch function */
 371             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 372             felec            = _mm_msub_pd( felec,sw , _mm_mul_pd(rinv00,_mm_mul_pd(velec,dsw)) );
 373             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 374             velec            = _mm_mul_pd(velec,sw);
 375             vvdw             = _mm_mul_pd(vvdw,sw);
 376             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 377
 378             /* Update potential sum for this i atom from the interaction with this j atom. */
 379             velec            = _mm_and_pd(velec,cutoff_mask);
 380             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 381             velecsum         = _mm_add_pd(velecsum,velec);
 382             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 383             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 384             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 385
 386             fscal            = _mm_add_pd(felec,fvdw);
 387
 388             fscal            = _mm_and_pd(fscal,cutoff_mask);
 389
 390             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 391
 392             /* Update vectorial force */
 393             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 394             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 395             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 396
 397             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 398                                                    _mm_mul_pd(dx00,fscal),
 399                                                    _mm_mul_pd(dy00,fscal),
 400                                                    _mm_mul_pd(dz00,fscal));
 401
 402             }
 403
 404             /* Inner loop uses 86 flops */
 405         }
 406
 407         /* End of innermost loop */
 408
 409         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 410                                               f+i_coord_offset,fshift+i_shift_offset);
 411
 412         ggid                        = gid[iidx];
 413         /* Update potential energies */
 414         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 415         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 416
 417         /* Increment number of inner iterations */
 418         inneriter                  += j_index_end - j_index_start;
 419
 420         /* Outer loop uses 9 flops */
 421     }
 422
 423     /* Increment number of outer iterations */
 424     outeriter        += nri;
 425
 426     /* Update outer/inner flops */
 427
 428     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*86);
 429 }
 430 /*
 431  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwLJSw_GeomP1P1_F_avx_128_fma_double
 432  * Electrostatics interaction: Ewald
 433  * VdW interaction:            LennardJones
 434  * Geometry:                   Particle-Particle
 435  * Calculate force/pot:        Force
 436  */
 437 void
 438 nb_kernel_ElecEwSw_VdwLJSw_GeomP1P1_F_avx_128_fma_double
 439                     (t_nblist                    * gmx_restrict       nlist,
 440                      rvec                        * gmx_restrict          xx,
 441                      rvec                        * gmx_restrict          ff,
 442                      struct t_forcerec           * gmx_restrict          fr,
 443                      t_mdatoms                   * gmx_restrict     mdatoms,
 444                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 445                      t_nrnb                      * gmx_restrict        nrnb)
 446 {
 447     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 448      * just 0 for non-waters.
 449      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 450      * jnr indices corresponding to data put in the four positions in the SIMD register.
 451      */
 452     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 453     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 454     int              jnrA,jnrB;
 455     int              j_coord_offsetA,j_coord_offsetB;
 456     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 457     real             rcutoff_scalar;
 458     real             *shiftvec,*fshift,*x,*f;
 459     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 460     int              vdwioffset0;
 461     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 462     int              vdwjidx0A,vdwjidx0B;
 463     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 464     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 465     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 466     real             *charge;
 467     int              nvdwtype;
 468     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 469     int              *vdwtype;
 470     real             *vdwparam;
 471     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 472     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 473     __m128i          ewitab;
 474     __m128d          ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 475     real             *ewtab;
 476     __m128d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 477     real             rswitch_scalar,d_scalar;
 478     __m128d          dummy_mask,cutoff_mask;
 479     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 480     __m128d          one     = _mm_set1_pd(1.0);
 481     __m128d          two     = _mm_set1_pd(2.0);
 482     x                = xx[0];
 483     f                = ff[0];
 484
 485     nri              = nlist->nri;
 486     iinr             = nlist->iinr;
 487     jindex           = nlist->jindex;
 488     jjnr             = nlist->jjnr;
 489     shiftidx         = nlist->shift;
 490     gid              = nlist->gid;
 491     shiftvec         = fr->shift_vec[0];
 492     fshift           = fr->fshift[0];
 493     facel            = _mm_set1_pd(fr->ic->epsfac);
 494     charge           = mdatoms->chargeA;
 495     nvdwtype         = fr->ntype;
 496     vdwparam         = fr->nbfp;
 497     vdwtype          = mdatoms->typeA;
 498
 499     sh_ewald         = _mm_set1_pd(fr->ic->sh_ewald);
 500     ewtab            = fr->ic->tabq_coul_FDV0;
 501     ewtabscale       = _mm_set1_pd(fr->ic->tabq_scale);
 502     ewtabhalfspace   = _mm_set1_pd(0.5/fr->ic->tabq_scale);
 503
 504     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 505     rcutoff_scalar   = fr->ic->rcoulomb;
 506     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 507     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 508
 509     rswitch_scalar   = fr->ic->rcoulomb_switch;
 510     rswitch          = _mm_set1_pd(rswitch_scalar);
 511     /* Setup switch parameters */
 512     d_scalar         = rcutoff_scalar-rswitch_scalar;
 513     d                = _mm_set1_pd(d_scalar);
 514     swV3             = _mm_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 515     swV4             = _mm_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 516     swV5             = _mm_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 517     swF2             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 518     swF3             = _mm_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 519     swF4             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 520
 521     /* Avoid stupid compiler warnings */
 522     jnrA = jnrB = 0;
 523     j_coord_offsetA = 0;
 524     j_coord_offsetB = 0;
 525
 526     outeriter        = 0;
 527     inneriter        = 0;
 528
 529     /* Start outer loop over neighborlists */
 530     for(iidx=0; iidx<nri; iidx++)
 531     {
 532         /* Load shift vector for this list */
 533         i_shift_offset   = DIM*shiftidx[iidx];
 534
 535         /* Load limits for loop over neighbors */
 536         j_index_start    = jindex[iidx];
 537         j_index_end      = jindex[iidx+1];
 538
 539         /* Get outer coordinate index */
 540         inr              = iinr[iidx];
 541         i_coord_offset   = DIM*inr;
 542
 543         /* Load i particle coords and add shift vector */
 544         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 545
 546         fix0             = _mm_setzero_pd();
 547         fiy0             = _mm_setzero_pd();
 548         fiz0             = _mm_setzero_pd();
 549
 550         /* Load parameters for i particles */
 551         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 552         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 553
 554         /* Start inner kernel loop */
 555         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 556         {
 557
 558             /* Get j neighbor index, and coordinate index */
 559             jnrA             = jjnr[jidx];
 560             jnrB             = jjnr[jidx+1];
 561             j_coord_offsetA  = DIM*jnrA;
 562             j_coord_offsetB  = DIM*jnrB;
 563
 564             /* load j atom coordinates */
 565             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 566                                               &jx0,&jy0,&jz0);
 567
 568             /* Calculate displacement vector */
 569             dx00             = _mm_sub_pd(ix0,jx0);
 570             dy00             = _mm_sub_pd(iy0,jy0);
 571             dz00             = _mm_sub_pd(iz0,jz0);
 572
 573             /* Calculate squared distance and things based on it */
 574             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 575
 576             rinv00           = avx128fma_invsqrt_d(rsq00);
 577
 578             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 579
 580             /* Load parameters for j particles */
 581             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 582             vdwjidx0A        = 2*vdwtype[jnrA+0];
 583             vdwjidx0B        = 2*vdwtype[jnrB+0];
 584
 585             /**************************
 586              * CALCULATE INTERACTIONS *
 587              **************************/
 588
 589             if (gmx_mm_any_lt(rsq00,rcutoff2))
 590             {
 591
 592             r00              = _mm_mul_pd(rsq00,rinv00);
 593
 594             /* Compute parameters for interactions between i and j atoms */
 595             qq00             = _mm_mul_pd(iq0,jq0);
 596             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 597                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 598
 599             /* EWALD ELECTROSTATICS */
 600
 601             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 602             ewrt             = _mm_mul_pd(r00,ewtabscale);
 603             ewitab           = _mm_cvttpd_epi32(ewrt);
 604 #ifdef __XOP__
 605             eweps            = _mm_frcz_pd(ewrt);
 606 #else
 607             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 608 #endif
 609             twoeweps         = _mm_add_pd(eweps,eweps);
 610             ewitab           = _mm_slli_epi32(ewitab,2);
 611             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 612             ewtabD           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 613             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 614             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 615             ewtabFn          = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,1) +2);
 616             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 617             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 618             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 619             velec            = _mm_mul_pd(qq00,_mm_sub_pd(rinv00,velec));
 620             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 621
 622             /* LENNARD-JONES DISPERSION/REPULSION */
 623
 624             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 625             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 626             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 627             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 628             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 629
 630             d                = _mm_sub_pd(r00,rswitch);
 631             d                = _mm_max_pd(d,_mm_setzero_pd());
 632             d2               = _mm_mul_pd(d,d);
 633             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 634
 635             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 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            = _mm_msub_pd( felec,sw , _mm_mul_pd(rinv00,_mm_mul_pd(velec,dsw)) );
 640             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 641             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 642
 643             fscal            = _mm_add_pd(felec,fvdw);
 644
 645             fscal            = _mm_and_pd(fscal,cutoff_mask);
 646
 647             /* Update vectorial force */
 648             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 649             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 650             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 651
 652             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 653                                                    _mm_mul_pd(dx00,fscal),
 654                                                    _mm_mul_pd(dy00,fscal),
 655                                                    _mm_mul_pd(dz00,fscal));
 656
 657             }
 658
 659             /* Inner loop uses 80 flops */
 660         }
 661
 662         if(jidx<j_index_end)
 663         {
 664
 665             jnrA             = jjnr[jidx];
 666             j_coord_offsetA  = DIM*jnrA;
 667
 668             /* load j atom coordinates */
 669             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 670                                               &jx0,&jy0,&jz0);
 671
 672             /* Calculate displacement vector */
 673             dx00             = _mm_sub_pd(ix0,jx0);
 674             dy00             = _mm_sub_pd(iy0,jy0);
 675             dz00             = _mm_sub_pd(iz0,jz0);
 676
 677             /* Calculate squared distance and things based on it */
 678             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 679
 680             rinv00           = avx128fma_invsqrt_d(rsq00);
 681
 682             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 683
 684             /* Load parameters for j particles */
 685             jq0              = _mm_load_sd(charge+jnrA+0);
 686             vdwjidx0A        = 2*vdwtype[jnrA+0];
 687
 688             /**************************
 689              * CALCULATE INTERACTIONS *
 690              **************************/
 691
 692             if (gmx_mm_any_lt(rsq00,rcutoff2))
 693             {
 694
 695             r00              = _mm_mul_pd(rsq00,rinv00);
 696
 697             /* Compute parameters for interactions between i and j atoms */
 698             qq00             = _mm_mul_pd(iq0,jq0);
 699             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 700
 701             /* EWALD ELECTROSTATICS */
 702
 703             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 704             ewrt             = _mm_mul_pd(r00,ewtabscale);
 705             ewitab           = _mm_cvttpd_epi32(ewrt);
 706 #ifdef __XOP__
 707             eweps            = _mm_frcz_pd(ewrt);
 708 #else
 709             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 710 #endif
 711             twoeweps         = _mm_add_pd(eweps,eweps);
 712             ewitab           = _mm_slli_epi32(ewitab,2);
 713             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 714             ewtabD           = _mm_setzero_pd();
 715             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 716             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 717             ewtabFn          = _mm_setzero_pd();
 718             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 719             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 720             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 721             velec            = _mm_mul_pd(qq00,_mm_sub_pd(rinv00,velec));
 722             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 723
 724             /* LENNARD-JONES DISPERSION/REPULSION */
 725
 726             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 727             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 728             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 729             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 730             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 731
 732             d                = _mm_sub_pd(r00,rswitch);
 733             d                = _mm_max_pd(d,_mm_setzero_pd());
 734             d2               = _mm_mul_pd(d,d);
 735             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 736
 737             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 738
 739             /* Evaluate switch function */
 740             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 741             felec            = _mm_msub_pd( felec,sw , _mm_mul_pd(rinv00,_mm_mul_pd(velec,dsw)) );
 742             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 743             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 744
 745             fscal            = _mm_add_pd(felec,fvdw);
 746
 747             fscal            = _mm_and_pd(fscal,cutoff_mask);
 748
 749             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 750
 751             /* Update vectorial force */
 752             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 753             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 754             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 755
 756             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 757                                                    _mm_mul_pd(dx00,fscal),
 758                                                    _mm_mul_pd(dy00,fscal),
 759                                                    _mm_mul_pd(dz00,fscal));
 760
 761             }
 762
 763             /* Inner loop uses 80 flops */
 764         }
 765
 766         /* End of innermost loop */
 767
 768         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 769                                               f+i_coord_offset,fshift+i_shift_offset);
 770
 771         /* Increment number of inner iterations */
 772         inneriter                  += j_index_end - j_index_start;
 773
 774         /* Outer loop uses 7 flops */
 775     }
 776
 777     /* Increment number of outer iterations */
 778     outeriter        += nri;
 779
 780     /* Update outer/inner flops */
 781
 782     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*80);
 783 }