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) 2012,2013,2014, 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_128_fma_double kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_double.h"
  48 #include "kernelutil_x86_avx_128_fma_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwLJSw_GeomP1P1_VF_avx_128_fma_double
  52  * Electrostatics interaction: Ewald
  53  * VdW interaction:            LennardJones
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecEwSw_VdwLJSw_GeomP1P1_VF_avx_128_fma_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 refer to j loop unrolling done with SSE double precision, e.g. for the two 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;
  75     int              j_coord_offsetA,j_coord_offsetB;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  80     int              vdwioffset0;
  81     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  82     int              vdwjidx0A,vdwjidx0B;
  83     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  84     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  85     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  86     real             *charge;
  87     int              nvdwtype;
  88     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  89     int              *vdwtype;
  90     real             *vdwparam;
  91     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
  92     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
  93     __m128i          ewitab;
  94     __m128d          ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  95     real             *ewtab;
  96     __m128d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  97     real             rswitch_scalar,d_scalar;
  98     __m128d          dummy_mask,cutoff_mask;
  99     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 100     __m128d          one     = _mm_set1_pd(1.0);
 101     __m128d          two     = _mm_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            = _mm_set1_pd(fr->epsfac);
 114     charge           = mdatoms->chargeA;
 115     nvdwtype         = fr->ntype;
 116     vdwparam         = fr->nbfp;
 117     vdwtype          = mdatoms->typeA;
 118
 119     sh_ewald         = _mm_set1_pd(fr->ic->sh_ewald);
 120     ewtab            = fr->ic->tabq_coul_FDV0;
 121     ewtabscale       = _mm_set1_pd(fr->ic->tabq_scale);
 122     ewtabhalfspace   = _mm_set1_pd(0.5/fr->ic->tabq_scale);
 123
 124     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 125     rcutoff_scalar   = fr->rcoulomb;
 126     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 127     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 128
 129     rswitch_scalar   = fr->rcoulomb_switch;
 130     rswitch          = _mm_set1_pd(rswitch_scalar);
 131     /* Setup switch parameters */
 132     d_scalar         = rcutoff_scalar-rswitch_scalar;
 133     d                = _mm_set1_pd(d_scalar);
 134     swV3             = _mm_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 135     swV4             = _mm_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 136     swV5             = _mm_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 137     swF2             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 138     swF3             = _mm_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 139     swF4             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 140
 141     /* Avoid stupid compiler warnings */
 142     jnrA = jnrB = 0;
 143     j_coord_offsetA = 0;
 144     j_coord_offsetB = 0;
 145
 146     outeriter        = 0;
 147     inneriter        = 0;
 148
 149     /* Start outer loop over neighborlists */
 150     for(iidx=0; iidx<nri; iidx++)
 151     {
 152         /* Load shift vector for this list */
 153         i_shift_offset   = DIM*shiftidx[iidx];
 154
 155         /* Load limits for loop over neighbors */
 156         j_index_start    = jindex[iidx];
 157         j_index_end      = jindex[iidx+1];
 158
 159         /* Get outer coordinate index */
 160         inr              = iinr[iidx];
 161         i_coord_offset   = DIM*inr;
 162
 163         /* Load i particle coords and add shift vector */
 164         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 165
 166         fix0             = _mm_setzero_pd();
 167         fiy0             = _mm_setzero_pd();
 168         fiz0             = _mm_setzero_pd();
 169
 170         /* Load parameters for i particles */
 171         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 172         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 173
 174         /* Reset potential sums */
 175         velecsum         = _mm_setzero_pd();
 176         vvdwsum          = _mm_setzero_pd();
 177
 178         /* Start inner kernel loop */
 179         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 180         {
 181
 182             /* Get j neighbor index, and coordinate index */
 183             jnrA             = jjnr[jidx];
 184             jnrB             = jjnr[jidx+1];
 185             j_coord_offsetA  = DIM*jnrA;
 186             j_coord_offsetB  = DIM*jnrB;
 187
 188             /* load j atom coordinates */
 189             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 190                                               &jx0,&jy0,&jz0);
 191
 192             /* Calculate displacement vector */
 193             dx00             = _mm_sub_pd(ix0,jx0);
 194             dy00             = _mm_sub_pd(iy0,jy0);
 195             dz00             = _mm_sub_pd(iz0,jz0);
 196
 197             /* Calculate squared distance and things based on it */
 198             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 199
 200             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 201
 202             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 203
 204             /* Load parameters for j particles */
 205             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 206             vdwjidx0A        = 2*vdwtype[jnrA+0];
 207             vdwjidx0B        = 2*vdwtype[jnrB+0];
 208
 209             /**************************
 210              * CALCULATE INTERACTIONS *
 211              **************************/
 212
 213             if (gmx_mm_any_lt(rsq00,rcutoff2))
 214             {
 215
 216             r00              = _mm_mul_pd(rsq00,rinv00);
 217
 218             /* Compute parameters for interactions between i and j atoms */
 219             qq00             = _mm_mul_pd(iq0,jq0);
 220             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 221                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 222
 223             /* EWALD ELECTROSTATICS */
 224
 225             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 226             ewrt             = _mm_mul_pd(r00,ewtabscale);
 227             ewitab           = _mm_cvttpd_epi32(ewrt);
 228 #ifdef __XOP__
 229             eweps            = _mm_frcz_pd(ewrt);
 230 #else
 231             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 232 #endif
 233             twoeweps         = _mm_add_pd(eweps,eweps);
 234             ewitab           = _mm_slli_epi32(ewitab,2);
 235             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 236             ewtabD           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 237             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 238             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 239             ewtabFn          = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,1) +2);
 240             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 241             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 242             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 243             velec            = _mm_mul_pd(qq00,_mm_sub_pd(rinv00,velec));
 244             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 245
 246             /* LENNARD-JONES DISPERSION/REPULSION */
 247
 248             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 249             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 250             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 251             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 252             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 253
 254             d                = _mm_sub_pd(r00,rswitch);
 255             d                = _mm_max_pd(d,_mm_setzero_pd());
 256             d2               = _mm_mul_pd(d,d);
 257             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 258
 259             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 260
 261             /* Evaluate switch function */
 262             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 263             felec            = _mm_msub_pd( felec,sw , _mm_mul_pd(rinv00,_mm_mul_pd(velec,dsw)) );
 264             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 265             velec            = _mm_mul_pd(velec,sw);
 266             vvdw             = _mm_mul_pd(vvdw,sw);
 267             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 268
 269             /* Update potential sum for this i atom from the interaction with this j atom. */
 270             velec            = _mm_and_pd(velec,cutoff_mask);
 271             velecsum         = _mm_add_pd(velecsum,velec);
 272             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 273             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 274
 275             fscal            = _mm_add_pd(felec,fvdw);
 276
 277             fscal            = _mm_and_pd(fscal,cutoff_mask);
 278
 279             /* Update vectorial force */
 280             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 281             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 282             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 283
 284             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 285                                                    _mm_mul_pd(dx00,fscal),
 286                                                    _mm_mul_pd(dy00,fscal),
 287                                                    _mm_mul_pd(dz00,fscal));
 288
 289             }
 290
 291             /* Inner loop uses 86 flops */
 292         }
 293
 294         if(jidx<j_index_end)
 295         {
 296
 297             jnrA             = jjnr[jidx];
 298             j_coord_offsetA  = DIM*jnrA;
 299
 300             /* load j atom coordinates */
 301             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 302                                               &jx0,&jy0,&jz0);
 303
 304             /* Calculate displacement vector */
 305             dx00             = _mm_sub_pd(ix0,jx0);
 306             dy00             = _mm_sub_pd(iy0,jy0);
 307             dz00             = _mm_sub_pd(iz0,jz0);
 308
 309             /* Calculate squared distance and things based on it */
 310             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 311
 312             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 313
 314             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 315
 316             /* Load parameters for j particles */
 317             jq0              = _mm_load_sd(charge+jnrA+0);
 318             vdwjidx0A        = 2*vdwtype[jnrA+0];
 319
 320             /**************************
 321              * CALCULATE INTERACTIONS *
 322              **************************/
 323
 324             if (gmx_mm_any_lt(rsq00,rcutoff2))
 325             {
 326
 327             r00              = _mm_mul_pd(rsq00,rinv00);
 328
 329             /* Compute parameters for interactions between i and j atoms */
 330             qq00             = _mm_mul_pd(iq0,jq0);
 331             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 332
 333             /* EWALD ELECTROSTATICS */
 334
 335             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 336             ewrt             = _mm_mul_pd(r00,ewtabscale);
 337             ewitab           = _mm_cvttpd_epi32(ewrt);
 338 #ifdef __XOP__
 339             eweps            = _mm_frcz_pd(ewrt);
 340 #else
 341             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 342 #endif
 343             twoeweps         = _mm_add_pd(eweps,eweps);
 344             ewitab           = _mm_slli_epi32(ewitab,2);
 345             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 346             ewtabD           = _mm_setzero_pd();
 347             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 348             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 349             ewtabFn          = _mm_setzero_pd();
 350             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 351             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 352             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 353             velec            = _mm_mul_pd(qq00,_mm_sub_pd(rinv00,velec));
 354             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 355
 356             /* LENNARD-JONES DISPERSION/REPULSION */
 357
 358             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 359             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 360             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 361             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 362             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 363
 364             d                = _mm_sub_pd(r00,rswitch);
 365             d                = _mm_max_pd(d,_mm_setzero_pd());
 366             d2               = _mm_mul_pd(d,d);
 367             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 368
 369             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 370
 371             /* Evaluate switch function */
 372             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 373             felec            = _mm_msub_pd( felec,sw , _mm_mul_pd(rinv00,_mm_mul_pd(velec,dsw)) );
 374             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 375             velec            = _mm_mul_pd(velec,sw);
 376             vvdw             = _mm_mul_pd(vvdw,sw);
 377             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 378
 379             /* Update potential sum for this i atom from the interaction with this j atom. */
 380             velec            = _mm_and_pd(velec,cutoff_mask);
 381             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 382             velecsum         = _mm_add_pd(velecsum,velec);
 383             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 384             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 385             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 386
 387             fscal            = _mm_add_pd(felec,fvdw);
 388
 389             fscal            = _mm_and_pd(fscal,cutoff_mask);
 390
 391             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 392
 393             /* Update vectorial force */
 394             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 395             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 396             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 397
 398             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 399                                                    _mm_mul_pd(dx00,fscal),
 400                                                    _mm_mul_pd(dy00,fscal),
 401                                                    _mm_mul_pd(dz00,fscal));
 402
 403             }
 404
 405             /* Inner loop uses 86 flops */
 406         }
 407
 408         /* End of innermost loop */
 409
 410         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 411                                               f+i_coord_offset,fshift+i_shift_offset);
 412
 413         ggid                        = gid[iidx];
 414         /* Update potential energies */
 415         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 416         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 417
 418         /* Increment number of inner iterations */
 419         inneriter                  += j_index_end - j_index_start;
 420
 421         /* Outer loop uses 9 flops */
 422     }
 423
 424     /* Increment number of outer iterations */
 425     outeriter        += nri;
 426
 427     /* Update outer/inner flops */
 428
 429     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*86);
 430 }
 431 /*
 432  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwLJSw_GeomP1P1_F_avx_128_fma_double
 433  * Electrostatics interaction: Ewald
 434  * VdW interaction:            LennardJones
 435  * Geometry:                   Particle-Particle
 436  * Calculate force/pot:        Force
 437  */
 438 void
 439 nb_kernel_ElecEwSw_VdwLJSw_GeomP1P1_F_avx_128_fma_double
 440                     (t_nblist                    * gmx_restrict       nlist,
 441                      rvec                        * gmx_restrict          xx,
 442                      rvec                        * gmx_restrict          ff,
 443                      t_forcerec                  * gmx_restrict          fr,
 444                      t_mdatoms                   * gmx_restrict     mdatoms,
 445                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 446                      t_nrnb                      * gmx_restrict        nrnb)
 447 {
 448     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 449      * just 0 for non-waters.
 450      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 451      * jnr indices corresponding to data put in the four positions in the SIMD register.
 452      */
 453     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 454     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 455     int              jnrA,jnrB;
 456     int              j_coord_offsetA,j_coord_offsetB;
 457     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 458     real             rcutoff_scalar;
 459     real             *shiftvec,*fshift,*x,*f;
 460     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 461     int              vdwioffset0;
 462     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 463     int              vdwjidx0A,vdwjidx0B;
 464     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 465     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 466     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 467     real             *charge;
 468     int              nvdwtype;
 469     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 470     int              *vdwtype;
 471     real             *vdwparam;
 472     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 473     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 474     __m128i          ewitab;
 475     __m128d          ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 476     real             *ewtab;
 477     __m128d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 478     real             rswitch_scalar,d_scalar;
 479     __m128d          dummy_mask,cutoff_mask;
 480     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 481     __m128d          one     = _mm_set1_pd(1.0);
 482     __m128d          two     = _mm_set1_pd(2.0);
 483     x                = xx[0];
 484     f                = ff[0];
 485
 486     nri              = nlist->nri;
 487     iinr             = nlist->iinr;
 488     jindex           = nlist->jindex;
 489     jjnr             = nlist->jjnr;
 490     shiftidx         = nlist->shift;
 491     gid              = nlist->gid;
 492     shiftvec         = fr->shift_vec[0];
 493     fshift           = fr->fshift[0];
 494     facel            = _mm_set1_pd(fr->epsfac);
 495     charge           = mdatoms->chargeA;
 496     nvdwtype         = fr->ntype;
 497     vdwparam         = fr->nbfp;
 498     vdwtype          = mdatoms->typeA;
 499
 500     sh_ewald         = _mm_set1_pd(fr->ic->sh_ewald);
 501     ewtab            = fr->ic->tabq_coul_FDV0;
 502     ewtabscale       = _mm_set1_pd(fr->ic->tabq_scale);
 503     ewtabhalfspace   = _mm_set1_pd(0.5/fr->ic->tabq_scale);
 504
 505     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 506     rcutoff_scalar   = fr->rcoulomb;
 507     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 508     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 509
 510     rswitch_scalar   = fr->rcoulomb_switch;
 511     rswitch          = _mm_set1_pd(rswitch_scalar);
 512     /* Setup switch parameters */
 513     d_scalar         = rcutoff_scalar-rswitch_scalar;
 514     d                = _mm_set1_pd(d_scalar);
 515     swV3             = _mm_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 516     swV4             = _mm_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 517     swV5             = _mm_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 518     swF2             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 519     swF3             = _mm_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 520     swF4             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 521
 522     /* Avoid stupid compiler warnings */
 523     jnrA = jnrB = 0;
 524     j_coord_offsetA = 0;
 525     j_coord_offsetB = 0;
 526
 527     outeriter        = 0;
 528     inneriter        = 0;
 529
 530     /* Start outer loop over neighborlists */
 531     for(iidx=0; iidx<nri; iidx++)
 532     {
 533         /* Load shift vector for this list */
 534         i_shift_offset   = DIM*shiftidx[iidx];
 535
 536         /* Load limits for loop over neighbors */
 537         j_index_start    = jindex[iidx];
 538         j_index_end      = jindex[iidx+1];
 539
 540         /* Get outer coordinate index */
 541         inr              = iinr[iidx];
 542         i_coord_offset   = DIM*inr;
 543
 544         /* Load i particle coords and add shift vector */
 545         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 546
 547         fix0             = _mm_setzero_pd();
 548         fiy0             = _mm_setzero_pd();
 549         fiz0             = _mm_setzero_pd();
 550
 551         /* Load parameters for i particles */
 552         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 553         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 554
 555         /* Start inner kernel loop */
 556         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 557         {
 558
 559             /* Get j neighbor index, and coordinate index */
 560             jnrA             = jjnr[jidx];
 561             jnrB             = jjnr[jidx+1];
 562             j_coord_offsetA  = DIM*jnrA;
 563             j_coord_offsetB  = DIM*jnrB;
 564
 565             /* load j atom coordinates */
 566             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 567                                               &jx0,&jy0,&jz0);
 568
 569             /* Calculate displacement vector */
 570             dx00             = _mm_sub_pd(ix0,jx0);
 571             dy00             = _mm_sub_pd(iy0,jy0);
 572             dz00             = _mm_sub_pd(iz0,jz0);
 573
 574             /* Calculate squared distance and things based on it */
 575             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 576
 577             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 578
 579             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 580
 581             /* Load parameters for j particles */
 582             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 583             vdwjidx0A        = 2*vdwtype[jnrA+0];
 584             vdwjidx0B        = 2*vdwtype[jnrB+0];
 585
 586             /**************************
 587              * CALCULATE INTERACTIONS *
 588              **************************/
 589
 590             if (gmx_mm_any_lt(rsq00,rcutoff2))
 591             {
 592
 593             r00              = _mm_mul_pd(rsq00,rinv00);
 594
 595             /* Compute parameters for interactions between i and j atoms */
 596             qq00             = _mm_mul_pd(iq0,jq0);
 597             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 598                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 599
 600             /* EWALD ELECTROSTATICS */
 601
 602             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 603             ewrt             = _mm_mul_pd(r00,ewtabscale);
 604             ewitab           = _mm_cvttpd_epi32(ewrt);
 605 #ifdef __XOP__
 606             eweps            = _mm_frcz_pd(ewrt);
 607 #else
 608             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 609 #endif
 610             twoeweps         = _mm_add_pd(eweps,eweps);
 611             ewitab           = _mm_slli_epi32(ewitab,2);
 612             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 613             ewtabD           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 614             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 615             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 616             ewtabFn          = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,1) +2);
 617             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 618             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 619             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 620             velec            = _mm_mul_pd(qq00,_mm_sub_pd(rinv00,velec));
 621             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 622
 623             /* LENNARD-JONES DISPERSION/REPULSION */
 624
 625             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 626             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 627             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 628             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 629             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 630
 631             d                = _mm_sub_pd(r00,rswitch);
 632             d                = _mm_max_pd(d,_mm_setzero_pd());
 633             d2               = _mm_mul_pd(d,d);
 634             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 635
 636             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 637
 638             /* Evaluate switch function */
 639             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 640             felec            = _mm_msub_pd( felec,sw , _mm_mul_pd(rinv00,_mm_mul_pd(velec,dsw)) );
 641             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 642             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 643
 644             fscal            = _mm_add_pd(felec,fvdw);
 645
 646             fscal            = _mm_and_pd(fscal,cutoff_mask);
 647
 648             /* Update vectorial force */
 649             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 650             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 651             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 652
 653             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 654                                                    _mm_mul_pd(dx00,fscal),
 655                                                    _mm_mul_pd(dy00,fscal),
 656                                                    _mm_mul_pd(dz00,fscal));
 657
 658             }
 659
 660             /* Inner loop uses 80 flops */
 661         }
 662
 663         if(jidx<j_index_end)
 664         {
 665
 666             jnrA             = jjnr[jidx];
 667             j_coord_offsetA  = DIM*jnrA;
 668
 669             /* load j atom coordinates */
 670             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 671                                               &jx0,&jy0,&jz0);
 672
 673             /* Calculate displacement vector */
 674             dx00             = _mm_sub_pd(ix0,jx0);
 675             dy00             = _mm_sub_pd(iy0,jy0);
 676             dz00             = _mm_sub_pd(iz0,jz0);
 677
 678             /* Calculate squared distance and things based on it */
 679             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 680
 681             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 682
 683             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 684
 685             /* Load parameters for j particles */
 686             jq0              = _mm_load_sd(charge+jnrA+0);
 687             vdwjidx0A        = 2*vdwtype[jnrA+0];
 688
 689             /**************************
 690              * CALCULATE INTERACTIONS *
 691              **************************/
 692
 693             if (gmx_mm_any_lt(rsq00,rcutoff2))
 694             {
 695
 696             r00              = _mm_mul_pd(rsq00,rinv00);
 697
 698             /* Compute parameters for interactions between i and j atoms */
 699             qq00             = _mm_mul_pd(iq0,jq0);
 700             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 701
 702             /* EWALD ELECTROSTATICS */
 703
 704             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 705             ewrt             = _mm_mul_pd(r00,ewtabscale);
 706             ewitab           = _mm_cvttpd_epi32(ewrt);
 707 #ifdef __XOP__
 708             eweps            = _mm_frcz_pd(ewrt);
 709 #else
 710             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 711 #endif
 712             twoeweps         = _mm_add_pd(eweps,eweps);
 713             ewitab           = _mm_slli_epi32(ewitab,2);
 714             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 715             ewtabD           = _mm_setzero_pd();
 716             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 717             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 718             ewtabFn          = _mm_setzero_pd();
 719             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 720             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 721             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 722             velec            = _mm_mul_pd(qq00,_mm_sub_pd(rinv00,velec));
 723             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 724
 725             /* LENNARD-JONES DISPERSION/REPULSION */
 726
 727             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 728             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 729             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 730             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 731             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 732
 733             d                = _mm_sub_pd(r00,rswitch);
 734             d                = _mm_max_pd(d,_mm_setzero_pd());
 735             d2               = _mm_mul_pd(d,d);
 736             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 737
 738             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 739
 740             /* Evaluate switch function */
 741             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 742             felec            = _mm_msub_pd( felec,sw , _mm_mul_pd(rinv00,_mm_mul_pd(velec,dsw)) );
 743             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 744             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 745
 746             fscal            = _mm_add_pd(felec,fvdw);
 747
 748             fscal            = _mm_and_pd(fscal,cutoff_mask);
 749
 750             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 751
 752             /* Update vectorial force */
 753             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 754             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 755             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 756
 757             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 758                                                    _mm_mul_pd(dx00,fscal),
 759                                                    _mm_mul_pd(dy00,fscal),
 760                                                    _mm_mul_pd(dz00,fscal));
 761
 762             }
 763
 764             /* Inner loop uses 80 flops */
 765         }
 766
 767         /* End of innermost loop */
 768
 769         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 770                                               f+i_coord_offset,fshift+i_shift_offset);
 771
 772         /* Increment number of inner iterations */
 773         inneriter                  += j_index_end - j_index_start;
 774
 775         /* Outer loop uses 7 flops */
 776     }
 777
 778     /* Increment number of outer iterations */
 779     outeriter        += nri;
 780
 781     /* Update outer/inner flops */
 782
 783     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*80);
 784 }