src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_avx_128_fma_single.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_single 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_single.h"
  50 #include "kernelutil_x86_avx_128_fma_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_VF_avx_128_fma_single
  54  * Electrostatics interaction: ReactionField
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_VF_avx_128_fma_single
  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,C,D refer to j loop unrolling done with AVX_128, e.g. for the four 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,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  83     real             scratch[4*DIM];
  84     __m128           fscal,rcutoff,rcutoff2,jidxall;
  85     int              vdwioffset0;
  86     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  88     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  90     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     int              nvdwtype;
  93     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  94     int              *vdwtype;
  95     real             *vdwparam;
  96     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  97     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  98     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  99     real             rswitch_scalar,d_scalar;
 100     __m128           dummy_mask,cutoff_mask;
 101     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 102     __m128           one     = _mm_set1_ps(1.0);
 103     __m128           two     = _mm_set1_ps(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_ps(fr->epsfac);
 116     charge           = mdatoms->chargeA;
 117     krf              = _mm_set1_ps(fr->ic->k_rf);
 118     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 119     crf              = _mm_set1_ps(fr->ic->c_rf);
 120     nvdwtype         = fr->ntype;
 121     vdwparam         = fr->nbfp;
 122     vdwtype          = mdatoms->typeA;
 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_ps(rcutoff_scalar);
 127     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 128
 129     rswitch_scalar   = fr->rvdw_switch;
 130     rswitch          = _mm_set1_ps(rswitch_scalar);
 131     /* Setup switch parameters */
 132     d_scalar         = rcutoff_scalar-rswitch_scalar;
 133     d                = _mm_set1_ps(d_scalar);
 134     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 135     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 136     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 137     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 138     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 139     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 140
 141     /* Avoid stupid compiler warnings */
 142     jnrA = jnrB = jnrC = jnrD = 0;
 143     j_coord_offsetA = 0;
 144     j_coord_offsetB = 0;
 145     j_coord_offsetC = 0;
 146     j_coord_offsetD = 0;
 147
 148     outeriter        = 0;
 149     inneriter        = 0;
 150
 151     for(iidx=0;iidx<4*DIM;iidx++)
 152     {
 153         scratch[iidx] = 0.0;
 154     }
 155
 156     /* Start outer loop over neighborlists */
 157     for(iidx=0; iidx<nri; iidx++)
 158     {
 159         /* Load shift vector for this list */
 160         i_shift_offset   = DIM*shiftidx[iidx];
 161
 162         /* Load limits for loop over neighbors */
 163         j_index_start    = jindex[iidx];
 164         j_index_end      = jindex[iidx+1];
 165
 166         /* Get outer coordinate index */
 167         inr              = iinr[iidx];
 168         i_coord_offset   = DIM*inr;
 169
 170         /* Load i particle coords and add shift vector */
 171         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 172
 173         fix0             = _mm_setzero_ps();
 174         fiy0             = _mm_setzero_ps();
 175         fiz0             = _mm_setzero_ps();
 176
 177         /* Load parameters for i particles */
 178         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 179         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 180
 181         /* Reset potential sums */
 182         velecsum         = _mm_setzero_ps();
 183         vvdwsum          = _mm_setzero_ps();
 184
 185         /* Start inner kernel loop */
 186         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 187         {
 188
 189             /* Get j neighbor index, and coordinate index */
 190             jnrA             = jjnr[jidx];
 191             jnrB             = jjnr[jidx+1];
 192             jnrC             = jjnr[jidx+2];
 193             jnrD             = jjnr[jidx+3];
 194             j_coord_offsetA  = DIM*jnrA;
 195             j_coord_offsetB  = DIM*jnrB;
 196             j_coord_offsetC  = DIM*jnrC;
 197             j_coord_offsetD  = DIM*jnrD;
 198
 199             /* load j atom coordinates */
 200             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 201                                               x+j_coord_offsetC,x+j_coord_offsetD,
 202                                               &jx0,&jy0,&jz0);
 203
 204             /* Calculate displacement vector */
 205             dx00             = _mm_sub_ps(ix0,jx0);
 206             dy00             = _mm_sub_ps(iy0,jy0);
 207             dz00             = _mm_sub_ps(iz0,jz0);
 208
 209             /* Calculate squared distance and things based on it */
 210             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 211
 212             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 213
 214             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 215
 216             /* Load parameters for j particles */
 217             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 218                                                               charge+jnrC+0,charge+jnrD+0);
 219             vdwjidx0A        = 2*vdwtype[jnrA+0];
 220             vdwjidx0B        = 2*vdwtype[jnrB+0];
 221             vdwjidx0C        = 2*vdwtype[jnrC+0];
 222             vdwjidx0D        = 2*vdwtype[jnrD+0];
 223
 224             /**************************
 225              * CALCULATE INTERACTIONS *
 226              **************************/
 227
 228             if (gmx_mm_any_lt(rsq00,rcutoff2))
 229             {
 230
 231             r00              = _mm_mul_ps(rsq00,rinv00);
 232
 233             /* Compute parameters for interactions between i and j atoms */
 234             qq00             = _mm_mul_ps(iq0,jq0);
 235             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 236                                          vdwparam+vdwioffset0+vdwjidx0B,
 237                                          vdwparam+vdwioffset0+vdwjidx0C,
 238                                          vdwparam+vdwioffset0+vdwjidx0D,
 239                                          &c6_00,&c12_00);
 240
 241             /* REACTION-FIELD ELECTROSTATICS */
 242             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 243             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 244
 245             /* LENNARD-JONES DISPERSION/REPULSION */
 246
 247             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 248             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 249             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 250             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 251             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 252
 253             d                = _mm_sub_ps(r00,rswitch);
 254             d                = _mm_max_ps(d,_mm_setzero_ps());
 255             d2               = _mm_mul_ps(d,d);
 256             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 257
 258             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(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             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 263             vvdw             = _mm_mul_ps(vvdw,sw);
 264             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 265
 266             /* Update potential sum for this i atom from the interaction with this j atom. */
 267             velec            = _mm_and_ps(velec,cutoff_mask);
 268             velecsum         = _mm_add_ps(velecsum,velec);
 269             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 270             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 271
 272             fscal            = _mm_add_ps(felec,fvdw);
 273
 274             fscal            = _mm_and_ps(fscal,cutoff_mask);
 275
 276              /* Update vectorial force */
 277             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 278             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 279             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 280
 281             fjptrA             = f+j_coord_offsetA;
 282             fjptrB             = f+j_coord_offsetB;
 283             fjptrC             = f+j_coord_offsetC;
 284             fjptrD             = f+j_coord_offsetD;
 285             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 286                                                    _mm_mul_ps(dx00,fscal),
 287                                                    _mm_mul_ps(dy00,fscal),
 288                                                    _mm_mul_ps(dz00,fscal));
 289
 290             }
 291
 292             /* Inner loop uses 73 flops */
 293         }
 294
 295         if(jidx<j_index_end)
 296         {
 297
 298             /* Get j neighbor index, and coordinate index */
 299             jnrlistA         = jjnr[jidx];
 300             jnrlistB         = jjnr[jidx+1];
 301             jnrlistC         = jjnr[jidx+2];
 302             jnrlistD         = jjnr[jidx+3];
 303             /* Sign of each element will be negative for non-real atoms.
 304              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 305              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 306              */
 307             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 308             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 309             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 310             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 311             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 312             j_coord_offsetA  = DIM*jnrA;
 313             j_coord_offsetB  = DIM*jnrB;
 314             j_coord_offsetC  = DIM*jnrC;
 315             j_coord_offsetD  = DIM*jnrD;
 316
 317             /* load j atom coordinates */
 318             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 319                                               x+j_coord_offsetC,x+j_coord_offsetD,
 320                                               &jx0,&jy0,&jz0);
 321
 322             /* Calculate displacement vector */
 323             dx00             = _mm_sub_ps(ix0,jx0);
 324             dy00             = _mm_sub_ps(iy0,jy0);
 325             dz00             = _mm_sub_ps(iz0,jz0);
 326
 327             /* Calculate squared distance and things based on it */
 328             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 329
 330             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 331
 332             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 333
 334             /* Load parameters for j particles */
 335             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 336                                                               charge+jnrC+0,charge+jnrD+0);
 337             vdwjidx0A        = 2*vdwtype[jnrA+0];
 338             vdwjidx0B        = 2*vdwtype[jnrB+0];
 339             vdwjidx0C        = 2*vdwtype[jnrC+0];
 340             vdwjidx0D        = 2*vdwtype[jnrD+0];
 341
 342             /**************************
 343              * CALCULATE INTERACTIONS *
 344              **************************/
 345
 346             if (gmx_mm_any_lt(rsq00,rcutoff2))
 347             {
 348
 349             r00              = _mm_mul_ps(rsq00,rinv00);
 350             r00              = _mm_andnot_ps(dummy_mask,r00);
 351
 352             /* Compute parameters for interactions between i and j atoms */
 353             qq00             = _mm_mul_ps(iq0,jq0);
 354             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 355                                          vdwparam+vdwioffset0+vdwjidx0B,
 356                                          vdwparam+vdwioffset0+vdwjidx0C,
 357                                          vdwparam+vdwioffset0+vdwjidx0D,
 358                                          &c6_00,&c12_00);
 359
 360             /* REACTION-FIELD ELECTROSTATICS */
 361             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 362             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 363
 364             /* LENNARD-JONES DISPERSION/REPULSION */
 365
 366             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 367             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 368             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 369             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 370             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 371
 372             d                = _mm_sub_ps(r00,rswitch);
 373             d                = _mm_max_ps(d,_mm_setzero_ps());
 374             d2               = _mm_mul_ps(d,d);
 375             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 376
 377             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 378
 379             /* Evaluate switch function */
 380             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 381             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 382             vvdw             = _mm_mul_ps(vvdw,sw);
 383             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 384
 385             /* Update potential sum for this i atom from the interaction with this j atom. */
 386             velec            = _mm_and_ps(velec,cutoff_mask);
 387             velec            = _mm_andnot_ps(dummy_mask,velec);
 388             velecsum         = _mm_add_ps(velecsum,velec);
 389             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 390             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 391             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 392
 393             fscal            = _mm_add_ps(felec,fvdw);
 394
 395             fscal            = _mm_and_ps(fscal,cutoff_mask);
 396
 397             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 398
 399              /* Update vectorial force */
 400             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 401             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 402             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 403
 404             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 405             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 406             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 407             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 408             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 409                                                    _mm_mul_ps(dx00,fscal),
 410                                                    _mm_mul_ps(dy00,fscal),
 411                                                    _mm_mul_ps(dz00,fscal));
 412
 413             }
 414
 415             /* Inner loop uses 74 flops */
 416         }
 417
 418         /* End of innermost loop */
 419
 420         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 421                                               f+i_coord_offset,fshift+i_shift_offset);
 422
 423         ggid                        = gid[iidx];
 424         /* Update potential energies */
 425         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 426         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 427
 428         /* Increment number of inner iterations */
 429         inneriter                  += j_index_end - j_index_start;
 430
 431         /* Outer loop uses 9 flops */
 432     }
 433
 434     /* Increment number of outer iterations */
 435     outeriter        += nri;
 436
 437     /* Update outer/inner flops */
 438
 439     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*74);
 440 }
 441 /*
 442  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_F_avx_128_fma_single
 443  * Electrostatics interaction: ReactionField
 444  * VdW interaction:            LennardJones
 445  * Geometry:                   Particle-Particle
 446  * Calculate force/pot:        Force
 447  */
 448 void
 449 nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_F_avx_128_fma_single
 450                     (t_nblist                    * gmx_restrict       nlist,
 451                      rvec                        * gmx_restrict          xx,
 452                      rvec                        * gmx_restrict          ff,
 453                      t_forcerec                  * gmx_restrict          fr,
 454                      t_mdatoms                   * gmx_restrict     mdatoms,
 455                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 456                      t_nrnb                      * gmx_restrict        nrnb)
 457 {
 458     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 459      * just 0 for non-waters.
 460      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 461      * jnr indices corresponding to data put in the four positions in the SIMD register.
 462      */
 463     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 464     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 465     int              jnrA,jnrB,jnrC,jnrD;
 466     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 467     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 468     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 469     real             rcutoff_scalar;
 470     real             *shiftvec,*fshift,*x,*f;
 471     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 472     real             scratch[4*DIM];
 473     __m128           fscal,rcutoff,rcutoff2,jidxall;
 474     int              vdwioffset0;
 475     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 476     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 477     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 478     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 479     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 480     real             *charge;
 481     int              nvdwtype;
 482     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 483     int              *vdwtype;
 484     real             *vdwparam;
 485     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 486     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 487     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 488     real             rswitch_scalar,d_scalar;
 489     __m128           dummy_mask,cutoff_mask;
 490     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 491     __m128           one     = _mm_set1_ps(1.0);
 492     __m128           two     = _mm_set1_ps(2.0);
 493     x                = xx[0];
 494     f                = ff[0];
 495
 496     nri              = nlist->nri;
 497     iinr             = nlist->iinr;
 498     jindex           = nlist->jindex;
 499     jjnr             = nlist->jjnr;
 500     shiftidx         = nlist->shift;
 501     gid              = nlist->gid;
 502     shiftvec         = fr->shift_vec[0];
 503     fshift           = fr->fshift[0];
 504     facel            = _mm_set1_ps(fr->epsfac);
 505     charge           = mdatoms->chargeA;
 506     krf              = _mm_set1_ps(fr->ic->k_rf);
 507     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 508     crf              = _mm_set1_ps(fr->ic->c_rf);
 509     nvdwtype         = fr->ntype;
 510     vdwparam         = fr->nbfp;
 511     vdwtype          = mdatoms->typeA;
 512
 513     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 514     rcutoff_scalar   = fr->rcoulomb;
 515     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 516     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 517
 518     rswitch_scalar   = fr->rvdw_switch;
 519     rswitch          = _mm_set1_ps(rswitch_scalar);
 520     /* Setup switch parameters */
 521     d_scalar         = rcutoff_scalar-rswitch_scalar;
 522     d                = _mm_set1_ps(d_scalar);
 523     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 524     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 525     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 526     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 527     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 528     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 529
 530     /* Avoid stupid compiler warnings */
 531     jnrA = jnrB = jnrC = jnrD = 0;
 532     j_coord_offsetA = 0;
 533     j_coord_offsetB = 0;
 534     j_coord_offsetC = 0;
 535     j_coord_offsetD = 0;
 536
 537     outeriter        = 0;
 538     inneriter        = 0;
 539
 540     for(iidx=0;iidx<4*DIM;iidx++)
 541     {
 542         scratch[iidx] = 0.0;
 543     }
 544
 545     /* Start outer loop over neighborlists */
 546     for(iidx=0; iidx<nri; iidx++)
 547     {
 548         /* Load shift vector for this list */
 549         i_shift_offset   = DIM*shiftidx[iidx];
 550
 551         /* Load limits for loop over neighbors */
 552         j_index_start    = jindex[iidx];
 553         j_index_end      = jindex[iidx+1];
 554
 555         /* Get outer coordinate index */
 556         inr              = iinr[iidx];
 557         i_coord_offset   = DIM*inr;
 558
 559         /* Load i particle coords and add shift vector */
 560         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 561
 562         fix0             = _mm_setzero_ps();
 563         fiy0             = _mm_setzero_ps();
 564         fiz0             = _mm_setzero_ps();
 565
 566         /* Load parameters for i particles */
 567         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 568         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 569
 570         /* Start inner kernel loop */
 571         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 572         {
 573
 574             /* Get j neighbor index, and coordinate index */
 575             jnrA             = jjnr[jidx];
 576             jnrB             = jjnr[jidx+1];
 577             jnrC             = jjnr[jidx+2];
 578             jnrD             = jjnr[jidx+3];
 579             j_coord_offsetA  = DIM*jnrA;
 580             j_coord_offsetB  = DIM*jnrB;
 581             j_coord_offsetC  = DIM*jnrC;
 582             j_coord_offsetD  = DIM*jnrD;
 583
 584             /* load j atom coordinates */
 585             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 586                                               x+j_coord_offsetC,x+j_coord_offsetD,
 587                                               &jx0,&jy0,&jz0);
 588
 589             /* Calculate displacement vector */
 590             dx00             = _mm_sub_ps(ix0,jx0);
 591             dy00             = _mm_sub_ps(iy0,jy0);
 592             dz00             = _mm_sub_ps(iz0,jz0);
 593
 594             /* Calculate squared distance and things based on it */
 595             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 596
 597             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 598
 599             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 600
 601             /* Load parameters for j particles */
 602             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 603                                                               charge+jnrC+0,charge+jnrD+0);
 604             vdwjidx0A        = 2*vdwtype[jnrA+0];
 605             vdwjidx0B        = 2*vdwtype[jnrB+0];
 606             vdwjidx0C        = 2*vdwtype[jnrC+0];
 607             vdwjidx0D        = 2*vdwtype[jnrD+0];
 608
 609             /**************************
 610              * CALCULATE INTERACTIONS *
 611              **************************/
 612
 613             if (gmx_mm_any_lt(rsq00,rcutoff2))
 614             {
 615
 616             r00              = _mm_mul_ps(rsq00,rinv00);
 617
 618             /* Compute parameters for interactions between i and j atoms */
 619             qq00             = _mm_mul_ps(iq0,jq0);
 620             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 621                                          vdwparam+vdwioffset0+vdwjidx0B,
 622                                          vdwparam+vdwioffset0+vdwjidx0C,
 623                                          vdwparam+vdwioffset0+vdwjidx0D,
 624                                          &c6_00,&c12_00);
 625
 626             /* REACTION-FIELD ELECTROSTATICS */
 627             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 628
 629             /* LENNARD-JONES DISPERSION/REPULSION */
 630
 631             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 632             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 633             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 634             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 635             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 636
 637             d                = _mm_sub_ps(r00,rswitch);
 638             d                = _mm_max_ps(d,_mm_setzero_ps());
 639             d2               = _mm_mul_ps(d,d);
 640             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 641
 642             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 643
 644             /* Evaluate switch function */
 645             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 646             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 647             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 648
 649             fscal            = _mm_add_ps(felec,fvdw);
 650
 651             fscal            = _mm_and_ps(fscal,cutoff_mask);
 652
 653              /* Update vectorial force */
 654             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 655             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 656             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 657
 658             fjptrA             = f+j_coord_offsetA;
 659             fjptrB             = f+j_coord_offsetB;
 660             fjptrC             = f+j_coord_offsetC;
 661             fjptrD             = f+j_coord_offsetD;
 662             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 663                                                    _mm_mul_ps(dx00,fscal),
 664                                                    _mm_mul_ps(dy00,fscal),
 665                                                    _mm_mul_ps(dz00,fscal));
 666
 667             }
 668
 669             /* Inner loop uses 64 flops */
 670         }
 671
 672         if(jidx<j_index_end)
 673         {
 674
 675             /* Get j neighbor index, and coordinate index */
 676             jnrlistA         = jjnr[jidx];
 677             jnrlistB         = jjnr[jidx+1];
 678             jnrlistC         = jjnr[jidx+2];
 679             jnrlistD         = jjnr[jidx+3];
 680             /* Sign of each element will be negative for non-real atoms.
 681              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 682              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 683              */
 684             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 685             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 686             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 687             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 688             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 689             j_coord_offsetA  = DIM*jnrA;
 690             j_coord_offsetB  = DIM*jnrB;
 691             j_coord_offsetC  = DIM*jnrC;
 692             j_coord_offsetD  = DIM*jnrD;
 693
 694             /* load j atom coordinates */
 695             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 696                                               x+j_coord_offsetC,x+j_coord_offsetD,
 697                                               &jx0,&jy0,&jz0);
 698
 699             /* Calculate displacement vector */
 700             dx00             = _mm_sub_ps(ix0,jx0);
 701             dy00             = _mm_sub_ps(iy0,jy0);
 702             dz00             = _mm_sub_ps(iz0,jz0);
 703
 704             /* Calculate squared distance and things based on it */
 705             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 706
 707             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 708
 709             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 710
 711             /* Load parameters for j particles */
 712             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 713                                                               charge+jnrC+0,charge+jnrD+0);
 714             vdwjidx0A        = 2*vdwtype[jnrA+0];
 715             vdwjidx0B        = 2*vdwtype[jnrB+0];
 716             vdwjidx0C        = 2*vdwtype[jnrC+0];
 717             vdwjidx0D        = 2*vdwtype[jnrD+0];
 718
 719             /**************************
 720              * CALCULATE INTERACTIONS *
 721              **************************/
 722
 723             if (gmx_mm_any_lt(rsq00,rcutoff2))
 724             {
 725
 726             r00              = _mm_mul_ps(rsq00,rinv00);
 727             r00              = _mm_andnot_ps(dummy_mask,r00);
 728
 729             /* Compute parameters for interactions between i and j atoms */
 730             qq00             = _mm_mul_ps(iq0,jq0);
 731             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 732                                          vdwparam+vdwioffset0+vdwjidx0B,
 733                                          vdwparam+vdwioffset0+vdwjidx0C,
 734                                          vdwparam+vdwioffset0+vdwjidx0D,
 735                                          &c6_00,&c12_00);
 736
 737             /* REACTION-FIELD ELECTROSTATICS */
 738             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 739
 740             /* LENNARD-JONES DISPERSION/REPULSION */
 741
 742             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 743             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 744             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 745             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 746             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 747
 748             d                = _mm_sub_ps(r00,rswitch);
 749             d                = _mm_max_ps(d,_mm_setzero_ps());
 750             d2               = _mm_mul_ps(d,d);
 751             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 752
 753             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 754
 755             /* Evaluate switch function */
 756             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 757             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 758             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 759
 760             fscal            = _mm_add_ps(felec,fvdw);
 761
 762             fscal            = _mm_and_ps(fscal,cutoff_mask);
 763
 764             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 765
 766              /* Update vectorial force */
 767             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 768             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 769             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 770
 771             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 772             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 773             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 774             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 775             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 776                                                    _mm_mul_ps(dx00,fscal),
 777                                                    _mm_mul_ps(dy00,fscal),
 778                                                    _mm_mul_ps(dz00,fscal));
 779
 780             }
 781
 782             /* Inner loop uses 65 flops */
 783         }
 784
 785         /* End of innermost loop */
 786
 787         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 788                                               f+i_coord_offset,fshift+i_shift_offset);
 789
 790         /* Increment number of inner iterations */
 791         inneriter                  += j_index_end - j_index_start;
 792
 793         /* Outer loop uses 7 flops */
 794     }
 795
 796     /* Increment number of outer iterations */
 797     outeriter        += nri;
 798
 799     /* Update outer/inner flops */
 800
 801     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*65);
 802 }