src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecEwSw_VdwLJSw_GeomP1P1_avx_128_fma_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_double kernel generator.
  37  */
  38 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_avx_128_fma_double.h"
  50 #include "kernelutil_x86_avx_128_fma_double.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwLJSw_GeomP1P1_VF_avx_128_fma_double
  54  * Electrostatics interaction: Ewald
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecEwSw_VdwLJSw_GeomP1P1_VF_avx_128_fma_double
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB;
  77     int              j_coord_offsetA,j_coord_offsetB;
  78     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  79     real             rcutoff_scalar;
  80     real             *shiftvec,*fshift,*x,*f;
  81     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwjidx0A,vdwjidx0B;
  85     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  86     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  87     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  88     real             *charge;
  89     int              nvdwtype;
  90     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  91     int              *vdwtype;
  92     real             *vdwparam;
  93     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
  94     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
  95     __m128i          ewitab;
  96     __m128d          ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  97     real             *ewtab;
  98     __m128d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  99     real             rswitch_scalar,d_scalar;
 100     __m128d          dummy_mask,cutoff_mask;
 101     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 102     __m128d          one     = _mm_set1_pd(1.0);
 103     __m128d          two     = _mm_set1_pd(2.0);
 104     x                = xx[0];
 105     f                = ff[0];
 106
 107     nri              = nlist->nri;
 108     iinr             = nlist->iinr;
 109     jindex           = nlist->jindex;
 110     jjnr             = nlist->jjnr;
 111     shiftidx         = nlist->shift;
 112     gid              = nlist->gid;
 113     shiftvec         = fr->shift_vec[0];
 114     fshift           = fr->fshift[0];
 115     facel            = _mm_set1_pd(fr->epsfac);
 116     charge           = mdatoms->chargeA;
 117     nvdwtype         = fr->ntype;
 118     vdwparam         = fr->nbfp;
 119     vdwtype          = mdatoms->typeA;
 120
 121     sh_ewald         = _mm_set1_pd(fr->ic->sh_ewald);
 122     ewtab            = fr->ic->tabq_coul_FDV0;
 123     ewtabscale       = _mm_set1_pd(fr->ic->tabq_scale);
 124     ewtabhalfspace   = _mm_set1_pd(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->rcoulomb;
 128     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 129     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 130
 131     rswitch_scalar   = fr->rcoulomb_switch;
 132     rswitch          = _mm_set1_pd(rswitch_scalar);
 133     /* Setup switch parameters */
 134     d_scalar         = rcutoff_scalar-rswitch_scalar;
 135     d                = _mm_set1_pd(d_scalar);
 136     swV3             = _mm_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 137     swV4             = _mm_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 138     swV5             = _mm_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 139     swF2             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 140     swF3             = _mm_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 141     swF4             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 142
 143     /* Avoid stupid compiler warnings */
 144     jnrA = jnrB = 0;
 145     j_coord_offsetA = 0;
 146     j_coord_offsetB = 0;
 147
 148     outeriter        = 0;
 149     inneriter        = 0;
 150
 151     /* Start outer loop over neighborlists */
 152     for(iidx=0; iidx<nri; iidx++)
 153     {
 154         /* Load shift vector for this list */
 155         i_shift_offset   = DIM*shiftidx[iidx];
 156
 157         /* Load limits for loop over neighbors */
 158         j_index_start    = jindex[iidx];
 159         j_index_end      = jindex[iidx+1];
 160
 161         /* Get outer coordinate index */
 162         inr              = iinr[iidx];
 163         i_coord_offset   = DIM*inr;
 164
 165         /* Load i particle coords and add shift vector */
 166         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 167
 168         fix0             = _mm_setzero_pd();
 169         fiy0             = _mm_setzero_pd();
 170         fiz0             = _mm_setzero_pd();
 171
 172         /* Load parameters for i particles */
 173         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 174         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 175
 176         /* Reset potential sums */
 177         velecsum         = _mm_setzero_pd();
 178         vvdwsum          = _mm_setzero_pd();
 179
 180         /* Start inner kernel loop */
 181         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 182         {
 183
 184             /* Get j neighbor index, and coordinate index */
 185             jnrA             = jjnr[jidx];
 186             jnrB             = jjnr[jidx+1];
 187             j_coord_offsetA  = DIM*jnrA;
 188             j_coord_offsetB  = DIM*jnrB;
 189
 190             /* load j atom coordinates */
 191             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 192                                               &jx0,&jy0,&jz0);
 193
 194             /* Calculate displacement vector */
 195             dx00             = _mm_sub_pd(ix0,jx0);
 196             dy00             = _mm_sub_pd(iy0,jy0);
 197             dz00             = _mm_sub_pd(iz0,jz0);
 198
 199             /* Calculate squared distance and things based on it */
 200             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 201
 202             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 203
 204             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 205
 206             /* Load parameters for j particles */
 207             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 208             vdwjidx0A        = 2*vdwtype[jnrA+0];
 209             vdwjidx0B        = 2*vdwtype[jnrB+0];
 210
 211             /**************************
 212              * CALCULATE INTERACTIONS *
 213              **************************/
 214
 215             if (gmx_mm_any_lt(rsq00,rcutoff2))
 216             {
 217
 218             r00              = _mm_mul_pd(rsq00,rinv00);
 219
 220             /* Compute parameters for interactions between i and j atoms */
 221             qq00             = _mm_mul_pd(iq0,jq0);
 222             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 223                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 224
 225             /* EWALD ELECTROSTATICS */
 226
 227             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 228             ewrt             = _mm_mul_pd(r00,ewtabscale);
 229             ewitab           = _mm_cvttpd_epi32(ewrt);
 230 #ifdef __XOP__
 231             eweps            = _mm_frcz_pd(ewrt);
 232 #else
 233             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 234 #endif
 235             twoeweps         = _mm_add_pd(eweps,eweps);
 236             ewitab           = _mm_slli_epi32(ewitab,2);
 237             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 238             ewtabD           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 239             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 240             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 241             ewtabFn          = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,1) +2);
 242             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 243             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 244             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 245             velec            = _mm_mul_pd(qq00,_mm_sub_pd(rinv00,velec));
 246             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 247
 248             /* LENNARD-JONES DISPERSION/REPULSION */
 249
 250             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 251             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 252             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 253             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 254             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 255
 256             d                = _mm_sub_pd(r00,rswitch);
 257             d                = _mm_max_pd(d,_mm_setzero_pd());
 258             d2               = _mm_mul_pd(d,d);
 259             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 260
 261             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 262
 263             /* Evaluate switch function */
 264             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 265             felec            = _mm_msub_pd( felec,sw , _mm_mul_pd(rinv00,_mm_mul_pd(velec,dsw)) );
 266             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 267             velec            = _mm_mul_pd(velec,sw);
 268             vvdw             = _mm_mul_pd(vvdw,sw);
 269             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 270
 271             /* Update potential sum for this i atom from the interaction with this j atom. */
 272             velec            = _mm_and_pd(velec,cutoff_mask);
 273             velecsum         = _mm_add_pd(velecsum,velec);
 274             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 275             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 276
 277             fscal            = _mm_add_pd(felec,fvdw);
 278
 279             fscal            = _mm_and_pd(fscal,cutoff_mask);
 280
 281             /* Update vectorial force */
 282             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 283             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 284             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 285
 286             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 287                                                    _mm_mul_pd(dx00,fscal),
 288                                                    _mm_mul_pd(dy00,fscal),
 289                                                    _mm_mul_pd(dz00,fscal));
 290
 291             }
 292
 293             /* Inner loop uses 86 flops */
 294         }
 295
 296         if(jidx<j_index_end)
 297         {
 298
 299             jnrA             = jjnr[jidx];
 300             j_coord_offsetA  = DIM*jnrA;
 301
 302             /* load j atom coordinates */
 303             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 304                                               &jx0,&jy0,&jz0);
 305
 306             /* Calculate displacement vector */
 307             dx00             = _mm_sub_pd(ix0,jx0);
 308             dy00             = _mm_sub_pd(iy0,jy0);
 309             dz00             = _mm_sub_pd(iz0,jz0);
 310
 311             /* Calculate squared distance and things based on it */
 312             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 313
 314             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 315
 316             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 317
 318             /* Load parameters for j particles */
 319             jq0              = _mm_load_sd(charge+jnrA+0);
 320             vdwjidx0A        = 2*vdwtype[jnrA+0];
 321
 322             /**************************
 323              * CALCULATE INTERACTIONS *
 324              **************************/
 325
 326             if (gmx_mm_any_lt(rsq00,rcutoff2))
 327             {
 328
 329             r00              = _mm_mul_pd(rsq00,rinv00);
 330
 331             /* Compute parameters for interactions between i and j atoms */
 332             qq00             = _mm_mul_pd(iq0,jq0);
 333             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 334
 335             /* EWALD ELECTROSTATICS */
 336
 337             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 338             ewrt             = _mm_mul_pd(r00,ewtabscale);
 339             ewitab           = _mm_cvttpd_epi32(ewrt);
 340 #ifdef __XOP__
 341             eweps            = _mm_frcz_pd(ewrt);
 342 #else
 343             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 344 #endif
 345             twoeweps         = _mm_add_pd(eweps,eweps);
 346             ewitab           = _mm_slli_epi32(ewitab,2);
 347             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 348             ewtabD           = _mm_setzero_pd();
 349             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 350             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 351             ewtabFn          = _mm_setzero_pd();
 352             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 353             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 354             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 355             velec            = _mm_mul_pd(qq00,_mm_sub_pd(rinv00,velec));
 356             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 357
 358             /* LENNARD-JONES DISPERSION/REPULSION */
 359
 360             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 361             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 362             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 363             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 364             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 365
 366             d                = _mm_sub_pd(r00,rswitch);
 367             d                = _mm_max_pd(d,_mm_setzero_pd());
 368             d2               = _mm_mul_pd(d,d);
 369             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 370
 371             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 372
 373             /* Evaluate switch function */
 374             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 375             felec            = _mm_msub_pd( felec,sw , _mm_mul_pd(rinv00,_mm_mul_pd(velec,dsw)) );
 376             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 377             velec            = _mm_mul_pd(velec,sw);
 378             vvdw             = _mm_mul_pd(vvdw,sw);
 379             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 380
 381             /* Update potential sum for this i atom from the interaction with this j atom. */
 382             velec            = _mm_and_pd(velec,cutoff_mask);
 383             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 384             velecsum         = _mm_add_pd(velecsum,velec);
 385             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 386             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 387             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 388
 389             fscal            = _mm_add_pd(felec,fvdw);
 390
 391             fscal            = _mm_and_pd(fscal,cutoff_mask);
 392
 393             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 394
 395             /* Update vectorial force */
 396             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 397             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 398             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 399
 400             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 401                                                    _mm_mul_pd(dx00,fscal),
 402                                                    _mm_mul_pd(dy00,fscal),
 403                                                    _mm_mul_pd(dz00,fscal));
 404
 405             }
 406
 407             /* Inner loop uses 86 flops */
 408         }
 409
 410         /* End of innermost loop */
 411
 412         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 413                                               f+i_coord_offset,fshift+i_shift_offset);
 414
 415         ggid                        = gid[iidx];
 416         /* Update potential energies */
 417         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 418         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 419
 420         /* Increment number of inner iterations */
 421         inneriter                  += j_index_end - j_index_start;
 422
 423         /* Outer loop uses 9 flops */
 424     }
 425
 426     /* Increment number of outer iterations */
 427     outeriter        += nri;
 428
 429     /* Update outer/inner flops */
 430
 431     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*86);
 432 }
 433 /*
 434  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwLJSw_GeomP1P1_F_avx_128_fma_double
 435  * Electrostatics interaction: Ewald
 436  * VdW interaction:            LennardJones
 437  * Geometry:                   Particle-Particle
 438  * Calculate force/pot:        Force
 439  */
 440 void
 441 nb_kernel_ElecEwSw_VdwLJSw_GeomP1P1_F_avx_128_fma_double
 442                     (t_nblist                    * gmx_restrict       nlist,
 443                      rvec                        * gmx_restrict          xx,
 444                      rvec                        * gmx_restrict          ff,
 445                      t_forcerec                  * gmx_restrict          fr,
 446                      t_mdatoms                   * gmx_restrict     mdatoms,
 447                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 448                      t_nrnb                      * gmx_restrict        nrnb)
 449 {
 450     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 451      * just 0 for non-waters.
 452      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 453      * jnr indices corresponding to data put in the four positions in the SIMD register.
 454      */
 455     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 456     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 457     int              jnrA,jnrB;
 458     int              j_coord_offsetA,j_coord_offsetB;
 459     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 460     real             rcutoff_scalar;
 461     real             *shiftvec,*fshift,*x,*f;
 462     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 463     int              vdwioffset0;
 464     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 465     int              vdwjidx0A,vdwjidx0B;
 466     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 467     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 468     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 469     real             *charge;
 470     int              nvdwtype;
 471     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 472     int              *vdwtype;
 473     real             *vdwparam;
 474     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 475     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 476     __m128i          ewitab;
 477     __m128d          ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 478     real             *ewtab;
 479     __m128d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 480     real             rswitch_scalar,d_scalar;
 481     __m128d          dummy_mask,cutoff_mask;
 482     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 483     __m128d          one     = _mm_set1_pd(1.0);
 484     __m128d          two     = _mm_set1_pd(2.0);
 485     x                = xx[0];
 486     f                = ff[0];
 487
 488     nri              = nlist->nri;
 489     iinr             = nlist->iinr;
 490     jindex           = nlist->jindex;
 491     jjnr             = nlist->jjnr;
 492     shiftidx         = nlist->shift;
 493     gid              = nlist->gid;
 494     shiftvec         = fr->shift_vec[0];
 495     fshift           = fr->fshift[0];
 496     facel            = _mm_set1_pd(fr->epsfac);
 497     charge           = mdatoms->chargeA;
 498     nvdwtype         = fr->ntype;
 499     vdwparam         = fr->nbfp;
 500     vdwtype          = mdatoms->typeA;
 501
 502     sh_ewald         = _mm_set1_pd(fr->ic->sh_ewald);
 503     ewtab            = fr->ic->tabq_coul_FDV0;
 504     ewtabscale       = _mm_set1_pd(fr->ic->tabq_scale);
 505     ewtabhalfspace   = _mm_set1_pd(0.5/fr->ic->tabq_scale);
 506
 507     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 508     rcutoff_scalar   = fr->rcoulomb;
 509     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 510     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 511
 512     rswitch_scalar   = fr->rcoulomb_switch;
 513     rswitch          = _mm_set1_pd(rswitch_scalar);
 514     /* Setup switch parameters */
 515     d_scalar         = rcutoff_scalar-rswitch_scalar;
 516     d                = _mm_set1_pd(d_scalar);
 517     swV3             = _mm_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 518     swV4             = _mm_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 519     swV5             = _mm_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 520     swF2             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 521     swF3             = _mm_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 522     swF4             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 523
 524     /* Avoid stupid compiler warnings */
 525     jnrA = jnrB = 0;
 526     j_coord_offsetA = 0;
 527     j_coord_offsetB = 0;
 528
 529     outeriter        = 0;
 530     inneriter        = 0;
 531
 532     /* Start outer loop over neighborlists */
 533     for(iidx=0; iidx<nri; iidx++)
 534     {
 535         /* Load shift vector for this list */
 536         i_shift_offset   = DIM*shiftidx[iidx];
 537
 538         /* Load limits for loop over neighbors */
 539         j_index_start    = jindex[iidx];
 540         j_index_end      = jindex[iidx+1];
 541
 542         /* Get outer coordinate index */
 543         inr              = iinr[iidx];
 544         i_coord_offset   = DIM*inr;
 545
 546         /* Load i particle coords and add shift vector */
 547         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 548
 549         fix0             = _mm_setzero_pd();
 550         fiy0             = _mm_setzero_pd();
 551         fiz0             = _mm_setzero_pd();
 552
 553         /* Load parameters for i particles */
 554         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 555         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 556
 557         /* Start inner kernel loop */
 558         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 559         {
 560
 561             /* Get j neighbor index, and coordinate index */
 562             jnrA             = jjnr[jidx];
 563             jnrB             = jjnr[jidx+1];
 564             j_coord_offsetA  = DIM*jnrA;
 565             j_coord_offsetB  = DIM*jnrB;
 566
 567             /* load j atom coordinates */
 568             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 569                                               &jx0,&jy0,&jz0);
 570
 571             /* Calculate displacement vector */
 572             dx00             = _mm_sub_pd(ix0,jx0);
 573             dy00             = _mm_sub_pd(iy0,jy0);
 574             dz00             = _mm_sub_pd(iz0,jz0);
 575
 576             /* Calculate squared distance and things based on it */
 577             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 578
 579             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 580
 581             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 582
 583             /* Load parameters for j particles */
 584             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 585             vdwjidx0A        = 2*vdwtype[jnrA+0];
 586             vdwjidx0B        = 2*vdwtype[jnrB+0];
 587
 588             /**************************
 589              * CALCULATE INTERACTIONS *
 590              **************************/
 591
 592             if (gmx_mm_any_lt(rsq00,rcutoff2))
 593             {
 594
 595             r00              = _mm_mul_pd(rsq00,rinv00);
 596
 597             /* Compute parameters for interactions between i and j atoms */
 598             qq00             = _mm_mul_pd(iq0,jq0);
 599             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 600                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 601
 602             /* EWALD ELECTROSTATICS */
 603
 604             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 605             ewrt             = _mm_mul_pd(r00,ewtabscale);
 606             ewitab           = _mm_cvttpd_epi32(ewrt);
 607 #ifdef __XOP__
 608             eweps            = _mm_frcz_pd(ewrt);
 609 #else
 610             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 611 #endif
 612             twoeweps         = _mm_add_pd(eweps,eweps);
 613             ewitab           = _mm_slli_epi32(ewitab,2);
 614             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 615             ewtabD           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 616             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 617             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 618             ewtabFn          = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,1) +2);
 619             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 620             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 621             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 622             velec            = _mm_mul_pd(qq00,_mm_sub_pd(rinv00,velec));
 623             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 624
 625             /* LENNARD-JONES DISPERSION/REPULSION */
 626
 627             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 628             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 629             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 630             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 631             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 632
 633             d                = _mm_sub_pd(r00,rswitch);
 634             d                = _mm_max_pd(d,_mm_setzero_pd());
 635             d2               = _mm_mul_pd(d,d);
 636             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 637
 638             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 639
 640             /* Evaluate switch function */
 641             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 642             felec            = _mm_msub_pd( felec,sw , _mm_mul_pd(rinv00,_mm_mul_pd(velec,dsw)) );
 643             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 644             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 645
 646             fscal            = _mm_add_pd(felec,fvdw);
 647
 648             fscal            = _mm_and_pd(fscal,cutoff_mask);
 649
 650             /* Update vectorial force */
 651             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 652             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 653             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 654
 655             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 656                                                    _mm_mul_pd(dx00,fscal),
 657                                                    _mm_mul_pd(dy00,fscal),
 658                                                    _mm_mul_pd(dz00,fscal));
 659
 660             }
 661
 662             /* Inner loop uses 80 flops */
 663         }
 664
 665         if(jidx<j_index_end)
 666         {
 667
 668             jnrA             = jjnr[jidx];
 669             j_coord_offsetA  = DIM*jnrA;
 670
 671             /* load j atom coordinates */
 672             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 673                                               &jx0,&jy0,&jz0);
 674
 675             /* Calculate displacement vector */
 676             dx00             = _mm_sub_pd(ix0,jx0);
 677             dy00             = _mm_sub_pd(iy0,jy0);
 678             dz00             = _mm_sub_pd(iz0,jz0);
 679
 680             /* Calculate squared distance and things based on it */
 681             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 682
 683             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 684
 685             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 686
 687             /* Load parameters for j particles */
 688             jq0              = _mm_load_sd(charge+jnrA+0);
 689             vdwjidx0A        = 2*vdwtype[jnrA+0];
 690
 691             /**************************
 692              * CALCULATE INTERACTIONS *
 693              **************************/
 694
 695             if (gmx_mm_any_lt(rsq00,rcutoff2))
 696             {
 697
 698             r00              = _mm_mul_pd(rsq00,rinv00);
 699
 700             /* Compute parameters for interactions between i and j atoms */
 701             qq00             = _mm_mul_pd(iq0,jq0);
 702             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 703
 704             /* EWALD ELECTROSTATICS */
 705
 706             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 707             ewrt             = _mm_mul_pd(r00,ewtabscale);
 708             ewitab           = _mm_cvttpd_epi32(ewrt);
 709 #ifdef __XOP__
 710             eweps            = _mm_frcz_pd(ewrt);
 711 #else
 712             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 713 #endif
 714             twoeweps         = _mm_add_pd(eweps,eweps);
 715             ewitab           = _mm_slli_epi32(ewitab,2);
 716             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 717             ewtabD           = _mm_setzero_pd();
 718             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 719             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 720             ewtabFn          = _mm_setzero_pd();
 721             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 722             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 723             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 724             velec            = _mm_mul_pd(qq00,_mm_sub_pd(rinv00,velec));
 725             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 726
 727             /* LENNARD-JONES DISPERSION/REPULSION */
 728
 729             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 730             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 731             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 732             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 733             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 734
 735             d                = _mm_sub_pd(r00,rswitch);
 736             d                = _mm_max_pd(d,_mm_setzero_pd());
 737             d2               = _mm_mul_pd(d,d);
 738             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 739
 740             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 741
 742             /* Evaluate switch function */
 743             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 744             felec            = _mm_msub_pd( felec,sw , _mm_mul_pd(rinv00,_mm_mul_pd(velec,dsw)) );
 745             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 746             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 747
 748             fscal            = _mm_add_pd(felec,fvdw);
 749
 750             fscal            = _mm_and_pd(fscal,cutoff_mask);
 751
 752             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 753
 754             /* Update vectorial force */
 755             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 756             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 757             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 758
 759             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 760                                                    _mm_mul_pd(dx00,fscal),
 761                                                    _mm_mul_pd(dy00,fscal),
 762                                                    _mm_mul_pd(dz00,fscal));
 763
 764             }
 765
 766             /* Inner loop uses 80 flops */
 767         }
 768
 769         /* End of innermost loop */
 770
 771         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 772                                               f+i_coord_offset,fshift+i_shift_offset);
 773
 774         /* Increment number of inner iterations */
 775         inneriter                  += j_index_end - j_index_start;
 776
 777         /* Outer loop uses 7 flops */
 778     }
 779
 780     /* Increment number of outer iterations */
 781     outeriter        += nri;
 782
 783     /* Update outer/inner flops */
 784
 785     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*80);
 786 }