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) 2012,2013,2014, 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_single kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  48 #include "kernelutil_x86_avx_128_fma_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_VF_avx_128_fma_single
  52  * Electrostatics interaction: ReactionField
  53  * VdW interaction:            LennardJones
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_VF_avx_128_fma_single
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m128           fscal,rcutoff,rcutoff2,jidxall;
  83     int              vdwioffset0;
  84     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  85     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  86     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  87     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  88     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  89     real             *charge;
  90     int              nvdwtype;
  91     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  92     int              *vdwtype;
  93     real             *vdwparam;
  94     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  95     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  96     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  97     real             rswitch_scalar,d_scalar;
  98     __m128           dummy_mask,cutoff_mask;
  99     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 100     __m128           one     = _mm_set1_ps(1.0);
 101     __m128           two     = _mm_set1_ps(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_ps(fr->epsfac);
 114     charge           = mdatoms->chargeA;
 115     krf              = _mm_set1_ps(fr->ic->k_rf);
 116     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 117     crf              = _mm_set1_ps(fr->ic->c_rf);
 118     nvdwtype         = fr->ntype;
 119     vdwparam         = fr->nbfp;
 120     vdwtype          = mdatoms->typeA;
 121
 122     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 123     rcutoff_scalar   = fr->rcoulomb;
 124     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 125     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 126
 127     rswitch_scalar   = fr->rvdw_switch;
 128     rswitch          = _mm_set1_ps(rswitch_scalar);
 129     /* Setup switch parameters */
 130     d_scalar         = rcutoff_scalar-rswitch_scalar;
 131     d                = _mm_set1_ps(d_scalar);
 132     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 133     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 134     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 135     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 136     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 137     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 138
 139     /* Avoid stupid compiler warnings */
 140     jnrA = jnrB = jnrC = jnrD = 0;
 141     j_coord_offsetA = 0;
 142     j_coord_offsetB = 0;
 143     j_coord_offsetC = 0;
 144     j_coord_offsetD = 0;
 145
 146     outeriter        = 0;
 147     inneriter        = 0;
 148
 149     for(iidx=0;iidx<4*DIM;iidx++)
 150     {
 151         scratch[iidx] = 0.0;
 152     }
 153
 154     /* Start outer loop over neighborlists */
 155     for(iidx=0; iidx<nri; iidx++)
 156     {
 157         /* Load shift vector for this list */
 158         i_shift_offset   = DIM*shiftidx[iidx];
 159
 160         /* Load limits for loop over neighbors */
 161         j_index_start    = jindex[iidx];
 162         j_index_end      = jindex[iidx+1];
 163
 164         /* Get outer coordinate index */
 165         inr              = iinr[iidx];
 166         i_coord_offset   = DIM*inr;
 167
 168         /* Load i particle coords and add shift vector */
 169         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 170
 171         fix0             = _mm_setzero_ps();
 172         fiy0             = _mm_setzero_ps();
 173         fiz0             = _mm_setzero_ps();
 174
 175         /* Load parameters for i particles */
 176         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 177         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 178
 179         /* Reset potential sums */
 180         velecsum         = _mm_setzero_ps();
 181         vvdwsum          = _mm_setzero_ps();
 182
 183         /* Start inner kernel loop */
 184         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 185         {
 186
 187             /* Get j neighbor index, and coordinate index */
 188             jnrA             = jjnr[jidx];
 189             jnrB             = jjnr[jidx+1];
 190             jnrC             = jjnr[jidx+2];
 191             jnrD             = jjnr[jidx+3];
 192             j_coord_offsetA  = DIM*jnrA;
 193             j_coord_offsetB  = DIM*jnrB;
 194             j_coord_offsetC  = DIM*jnrC;
 195             j_coord_offsetD  = DIM*jnrD;
 196
 197             /* load j atom coordinates */
 198             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 199                                               x+j_coord_offsetC,x+j_coord_offsetD,
 200                                               &jx0,&jy0,&jz0);
 201
 202             /* Calculate displacement vector */
 203             dx00             = _mm_sub_ps(ix0,jx0);
 204             dy00             = _mm_sub_ps(iy0,jy0);
 205             dz00             = _mm_sub_ps(iz0,jz0);
 206
 207             /* Calculate squared distance and things based on it */
 208             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 209
 210             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 211
 212             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 213
 214             /* Load parameters for j particles */
 215             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 216                                                               charge+jnrC+0,charge+jnrD+0);
 217             vdwjidx0A        = 2*vdwtype[jnrA+0];
 218             vdwjidx0B        = 2*vdwtype[jnrB+0];
 219             vdwjidx0C        = 2*vdwtype[jnrC+0];
 220             vdwjidx0D        = 2*vdwtype[jnrD+0];
 221
 222             /**************************
 223              * CALCULATE INTERACTIONS *
 224              **************************/
 225
 226             if (gmx_mm_any_lt(rsq00,rcutoff2))
 227             {
 228
 229             r00              = _mm_mul_ps(rsq00,rinv00);
 230
 231             /* Compute parameters for interactions between i and j atoms */
 232             qq00             = _mm_mul_ps(iq0,jq0);
 233             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 234                                          vdwparam+vdwioffset0+vdwjidx0B,
 235                                          vdwparam+vdwioffset0+vdwjidx0C,
 236                                          vdwparam+vdwioffset0+vdwjidx0D,
 237                                          &c6_00,&c12_00);
 238
 239             /* REACTION-FIELD ELECTROSTATICS */
 240             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 241             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 242
 243             /* LENNARD-JONES DISPERSION/REPULSION */
 244
 245             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 246             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 247             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 248             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 249             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 250
 251             d                = _mm_sub_ps(r00,rswitch);
 252             d                = _mm_max_ps(d,_mm_setzero_ps());
 253             d2               = _mm_mul_ps(d,d);
 254             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 255
 256             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 257
 258             /* Evaluate switch function */
 259             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 260             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 261             vvdw             = _mm_mul_ps(vvdw,sw);
 262             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 263
 264             /* Update potential sum for this i atom from the interaction with this j atom. */
 265             velec            = _mm_and_ps(velec,cutoff_mask);
 266             velecsum         = _mm_add_ps(velecsum,velec);
 267             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 268             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 269
 270             fscal            = _mm_add_ps(felec,fvdw);
 271
 272             fscal            = _mm_and_ps(fscal,cutoff_mask);
 273
 274              /* Update vectorial force */
 275             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 276             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 277             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 278
 279             fjptrA             = f+j_coord_offsetA;
 280             fjptrB             = f+j_coord_offsetB;
 281             fjptrC             = f+j_coord_offsetC;
 282             fjptrD             = f+j_coord_offsetD;
 283             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 284                                                    _mm_mul_ps(dx00,fscal),
 285                                                    _mm_mul_ps(dy00,fscal),
 286                                                    _mm_mul_ps(dz00,fscal));
 287
 288             }
 289
 290             /* Inner loop uses 73 flops */
 291         }
 292
 293         if(jidx<j_index_end)
 294         {
 295
 296             /* Get j neighbor index, and coordinate index */
 297             jnrlistA         = jjnr[jidx];
 298             jnrlistB         = jjnr[jidx+1];
 299             jnrlistC         = jjnr[jidx+2];
 300             jnrlistD         = jjnr[jidx+3];
 301             /* Sign of each element will be negative for non-real atoms.
 302              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 303              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 304              */
 305             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 306             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 307             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 308             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 309             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 310             j_coord_offsetA  = DIM*jnrA;
 311             j_coord_offsetB  = DIM*jnrB;
 312             j_coord_offsetC  = DIM*jnrC;
 313             j_coord_offsetD  = DIM*jnrD;
 314
 315             /* load j atom coordinates */
 316             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 317                                               x+j_coord_offsetC,x+j_coord_offsetD,
 318                                               &jx0,&jy0,&jz0);
 319
 320             /* Calculate displacement vector */
 321             dx00             = _mm_sub_ps(ix0,jx0);
 322             dy00             = _mm_sub_ps(iy0,jy0);
 323             dz00             = _mm_sub_ps(iz0,jz0);
 324
 325             /* Calculate squared distance and things based on it */
 326             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 327
 328             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 329
 330             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 331
 332             /* Load parameters for j particles */
 333             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 334                                                               charge+jnrC+0,charge+jnrD+0);
 335             vdwjidx0A        = 2*vdwtype[jnrA+0];
 336             vdwjidx0B        = 2*vdwtype[jnrB+0];
 337             vdwjidx0C        = 2*vdwtype[jnrC+0];
 338             vdwjidx0D        = 2*vdwtype[jnrD+0];
 339
 340             /**************************
 341              * CALCULATE INTERACTIONS *
 342              **************************/
 343
 344             if (gmx_mm_any_lt(rsq00,rcutoff2))
 345             {
 346
 347             r00              = _mm_mul_ps(rsq00,rinv00);
 348             r00              = _mm_andnot_ps(dummy_mask,r00);
 349
 350             /* Compute parameters for interactions between i and j atoms */
 351             qq00             = _mm_mul_ps(iq0,jq0);
 352             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 353                                          vdwparam+vdwioffset0+vdwjidx0B,
 354                                          vdwparam+vdwioffset0+vdwjidx0C,
 355                                          vdwparam+vdwioffset0+vdwjidx0D,
 356                                          &c6_00,&c12_00);
 357
 358             /* REACTION-FIELD ELECTROSTATICS */
 359             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 360             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 361
 362             /* LENNARD-JONES DISPERSION/REPULSION */
 363
 364             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 365             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 366             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 367             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 368             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 369
 370             d                = _mm_sub_ps(r00,rswitch);
 371             d                = _mm_max_ps(d,_mm_setzero_ps());
 372             d2               = _mm_mul_ps(d,d);
 373             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 374
 375             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 376
 377             /* Evaluate switch function */
 378             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 379             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 380             vvdw             = _mm_mul_ps(vvdw,sw);
 381             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 382
 383             /* Update potential sum for this i atom from the interaction with this j atom. */
 384             velec            = _mm_and_ps(velec,cutoff_mask);
 385             velec            = _mm_andnot_ps(dummy_mask,velec);
 386             velecsum         = _mm_add_ps(velecsum,velec);
 387             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 388             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 389             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 390
 391             fscal            = _mm_add_ps(felec,fvdw);
 392
 393             fscal            = _mm_and_ps(fscal,cutoff_mask);
 394
 395             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 396
 397              /* Update vectorial force */
 398             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 399             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 400             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 401
 402             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 403             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 404             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 405             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 406             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 407                                                    _mm_mul_ps(dx00,fscal),
 408                                                    _mm_mul_ps(dy00,fscal),
 409                                                    _mm_mul_ps(dz00,fscal));
 410
 411             }
 412
 413             /* Inner loop uses 74 flops */
 414         }
 415
 416         /* End of innermost loop */
 417
 418         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 419                                               f+i_coord_offset,fshift+i_shift_offset);
 420
 421         ggid                        = gid[iidx];
 422         /* Update potential energies */
 423         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 424         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 425
 426         /* Increment number of inner iterations */
 427         inneriter                  += j_index_end - j_index_start;
 428
 429         /* Outer loop uses 9 flops */
 430     }
 431
 432     /* Increment number of outer iterations */
 433     outeriter        += nri;
 434
 435     /* Update outer/inner flops */
 436
 437     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*74);
 438 }
 439 /*
 440  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_F_avx_128_fma_single
 441  * Electrostatics interaction: ReactionField
 442  * VdW interaction:            LennardJones
 443  * Geometry:                   Particle-Particle
 444  * Calculate force/pot:        Force
 445  */
 446 void
 447 nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_F_avx_128_fma_single
 448                     (t_nblist                    * gmx_restrict       nlist,
 449                      rvec                        * gmx_restrict          xx,
 450                      rvec                        * gmx_restrict          ff,
 451                      t_forcerec                  * gmx_restrict          fr,
 452                      t_mdatoms                   * gmx_restrict     mdatoms,
 453                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 454                      t_nrnb                      * gmx_restrict        nrnb)
 455 {
 456     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 457      * just 0 for non-waters.
 458      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 459      * jnr indices corresponding to data put in the four positions in the SIMD register.
 460      */
 461     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 462     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 463     int              jnrA,jnrB,jnrC,jnrD;
 464     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 465     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 466     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 467     real             rcutoff_scalar;
 468     real             *shiftvec,*fshift,*x,*f;
 469     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 470     real             scratch[4*DIM];
 471     __m128           fscal,rcutoff,rcutoff2,jidxall;
 472     int              vdwioffset0;
 473     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 474     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 475     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 476     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 477     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 478     real             *charge;
 479     int              nvdwtype;
 480     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 481     int              *vdwtype;
 482     real             *vdwparam;
 483     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 484     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 485     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 486     real             rswitch_scalar,d_scalar;
 487     __m128           dummy_mask,cutoff_mask;
 488     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 489     __m128           one     = _mm_set1_ps(1.0);
 490     __m128           two     = _mm_set1_ps(2.0);
 491     x                = xx[0];
 492     f                = ff[0];
 493
 494     nri              = nlist->nri;
 495     iinr             = nlist->iinr;
 496     jindex           = nlist->jindex;
 497     jjnr             = nlist->jjnr;
 498     shiftidx         = nlist->shift;
 499     gid              = nlist->gid;
 500     shiftvec         = fr->shift_vec[0];
 501     fshift           = fr->fshift[0];
 502     facel            = _mm_set1_ps(fr->epsfac);
 503     charge           = mdatoms->chargeA;
 504     krf              = _mm_set1_ps(fr->ic->k_rf);
 505     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 506     crf              = _mm_set1_ps(fr->ic->c_rf);
 507     nvdwtype         = fr->ntype;
 508     vdwparam         = fr->nbfp;
 509     vdwtype          = mdatoms->typeA;
 510
 511     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 512     rcutoff_scalar   = fr->rcoulomb;
 513     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 514     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 515
 516     rswitch_scalar   = fr->rvdw_switch;
 517     rswitch          = _mm_set1_ps(rswitch_scalar);
 518     /* Setup switch parameters */
 519     d_scalar         = rcutoff_scalar-rswitch_scalar;
 520     d                = _mm_set1_ps(d_scalar);
 521     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 522     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 523     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 524     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 525     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 526     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 527
 528     /* Avoid stupid compiler warnings */
 529     jnrA = jnrB = jnrC = jnrD = 0;
 530     j_coord_offsetA = 0;
 531     j_coord_offsetB = 0;
 532     j_coord_offsetC = 0;
 533     j_coord_offsetD = 0;
 534
 535     outeriter        = 0;
 536     inneriter        = 0;
 537
 538     for(iidx=0;iidx<4*DIM;iidx++)
 539     {
 540         scratch[iidx] = 0.0;
 541     }
 542
 543     /* Start outer loop over neighborlists */
 544     for(iidx=0; iidx<nri; iidx++)
 545     {
 546         /* Load shift vector for this list */
 547         i_shift_offset   = DIM*shiftidx[iidx];
 548
 549         /* Load limits for loop over neighbors */
 550         j_index_start    = jindex[iidx];
 551         j_index_end      = jindex[iidx+1];
 552
 553         /* Get outer coordinate index */
 554         inr              = iinr[iidx];
 555         i_coord_offset   = DIM*inr;
 556
 557         /* Load i particle coords and add shift vector */
 558         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 559
 560         fix0             = _mm_setzero_ps();
 561         fiy0             = _mm_setzero_ps();
 562         fiz0             = _mm_setzero_ps();
 563
 564         /* Load parameters for i particles */
 565         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 566         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 567
 568         /* Start inner kernel loop */
 569         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 570         {
 571
 572             /* Get j neighbor index, and coordinate index */
 573             jnrA             = jjnr[jidx];
 574             jnrB             = jjnr[jidx+1];
 575             jnrC             = jjnr[jidx+2];
 576             jnrD             = jjnr[jidx+3];
 577             j_coord_offsetA  = DIM*jnrA;
 578             j_coord_offsetB  = DIM*jnrB;
 579             j_coord_offsetC  = DIM*jnrC;
 580             j_coord_offsetD  = DIM*jnrD;
 581
 582             /* load j atom coordinates */
 583             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 584                                               x+j_coord_offsetC,x+j_coord_offsetD,
 585                                               &jx0,&jy0,&jz0);
 586
 587             /* Calculate displacement vector */
 588             dx00             = _mm_sub_ps(ix0,jx0);
 589             dy00             = _mm_sub_ps(iy0,jy0);
 590             dz00             = _mm_sub_ps(iz0,jz0);
 591
 592             /* Calculate squared distance and things based on it */
 593             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 594
 595             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 596
 597             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 598
 599             /* Load parameters for j particles */
 600             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 601                                                               charge+jnrC+0,charge+jnrD+0);
 602             vdwjidx0A        = 2*vdwtype[jnrA+0];
 603             vdwjidx0B        = 2*vdwtype[jnrB+0];
 604             vdwjidx0C        = 2*vdwtype[jnrC+0];
 605             vdwjidx0D        = 2*vdwtype[jnrD+0];
 606
 607             /**************************
 608              * CALCULATE INTERACTIONS *
 609              **************************/
 610
 611             if (gmx_mm_any_lt(rsq00,rcutoff2))
 612             {
 613
 614             r00              = _mm_mul_ps(rsq00,rinv00);
 615
 616             /* Compute parameters for interactions between i and j atoms */
 617             qq00             = _mm_mul_ps(iq0,jq0);
 618             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 619                                          vdwparam+vdwioffset0+vdwjidx0B,
 620                                          vdwparam+vdwioffset0+vdwjidx0C,
 621                                          vdwparam+vdwioffset0+vdwjidx0D,
 622                                          &c6_00,&c12_00);
 623
 624             /* REACTION-FIELD ELECTROSTATICS */
 625             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 626
 627             /* LENNARD-JONES DISPERSION/REPULSION */
 628
 629             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 630             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 631             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 632             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 633             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 634
 635             d                = _mm_sub_ps(r00,rswitch);
 636             d                = _mm_max_ps(d,_mm_setzero_ps());
 637             d2               = _mm_mul_ps(d,d);
 638             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 639
 640             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 641
 642             /* Evaluate switch function */
 643             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 644             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 645             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 646
 647             fscal            = _mm_add_ps(felec,fvdw);
 648
 649             fscal            = _mm_and_ps(fscal,cutoff_mask);
 650
 651              /* Update vectorial force */
 652             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 653             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 654             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 655
 656             fjptrA             = f+j_coord_offsetA;
 657             fjptrB             = f+j_coord_offsetB;
 658             fjptrC             = f+j_coord_offsetC;
 659             fjptrD             = f+j_coord_offsetD;
 660             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 661                                                    _mm_mul_ps(dx00,fscal),
 662                                                    _mm_mul_ps(dy00,fscal),
 663                                                    _mm_mul_ps(dz00,fscal));
 664
 665             }
 666
 667             /* Inner loop uses 64 flops */
 668         }
 669
 670         if(jidx<j_index_end)
 671         {
 672
 673             /* Get j neighbor index, and coordinate index */
 674             jnrlistA         = jjnr[jidx];
 675             jnrlistB         = jjnr[jidx+1];
 676             jnrlistC         = jjnr[jidx+2];
 677             jnrlistD         = jjnr[jidx+3];
 678             /* Sign of each element will be negative for non-real atoms.
 679              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 680              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 681              */
 682             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 683             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 684             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 685             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 686             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 687             j_coord_offsetA  = DIM*jnrA;
 688             j_coord_offsetB  = DIM*jnrB;
 689             j_coord_offsetC  = DIM*jnrC;
 690             j_coord_offsetD  = DIM*jnrD;
 691
 692             /* load j atom coordinates */
 693             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 694                                               x+j_coord_offsetC,x+j_coord_offsetD,
 695                                               &jx0,&jy0,&jz0);
 696
 697             /* Calculate displacement vector */
 698             dx00             = _mm_sub_ps(ix0,jx0);
 699             dy00             = _mm_sub_ps(iy0,jy0);
 700             dz00             = _mm_sub_ps(iz0,jz0);
 701
 702             /* Calculate squared distance and things based on it */
 703             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 704
 705             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 706
 707             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 708
 709             /* Load parameters for j particles */
 710             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 711                                                               charge+jnrC+0,charge+jnrD+0);
 712             vdwjidx0A        = 2*vdwtype[jnrA+0];
 713             vdwjidx0B        = 2*vdwtype[jnrB+0];
 714             vdwjidx0C        = 2*vdwtype[jnrC+0];
 715             vdwjidx0D        = 2*vdwtype[jnrD+0];
 716
 717             /**************************
 718              * CALCULATE INTERACTIONS *
 719              **************************/
 720
 721             if (gmx_mm_any_lt(rsq00,rcutoff2))
 722             {
 723
 724             r00              = _mm_mul_ps(rsq00,rinv00);
 725             r00              = _mm_andnot_ps(dummy_mask,r00);
 726
 727             /* Compute parameters for interactions between i and j atoms */
 728             qq00             = _mm_mul_ps(iq0,jq0);
 729             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 730                                          vdwparam+vdwioffset0+vdwjidx0B,
 731                                          vdwparam+vdwioffset0+vdwjidx0C,
 732                                          vdwparam+vdwioffset0+vdwjidx0D,
 733                                          &c6_00,&c12_00);
 734
 735             /* REACTION-FIELD ELECTROSTATICS */
 736             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 737
 738             /* LENNARD-JONES DISPERSION/REPULSION */
 739
 740             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 741             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 742             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 743             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 744             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 745
 746             d                = _mm_sub_ps(r00,rswitch);
 747             d                = _mm_max_ps(d,_mm_setzero_ps());
 748             d2               = _mm_mul_ps(d,d);
 749             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 750
 751             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 752
 753             /* Evaluate switch function */
 754             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 755             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 756             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 757
 758             fscal            = _mm_add_ps(felec,fvdw);
 759
 760             fscal            = _mm_and_ps(fscal,cutoff_mask);
 761
 762             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 763
 764              /* Update vectorial force */
 765             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 766             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 767             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 768
 769             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 770             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 771             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 772             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 773             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 774                                                    _mm_mul_ps(dx00,fscal),
 775                                                    _mm_mul_ps(dy00,fscal),
 776                                                    _mm_mul_ps(dz00,fscal));
 777
 778             }
 779
 780             /* Inner loop uses 65 flops */
 781         }
 782
 783         /* End of innermost loop */
 784
 785         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 786                                               f+i_coord_offset,fshift+i_shift_offset);
 787
 788         /* Increment number of inner iterations */
 789         inneriter                  += j_index_end - j_index_start;
 790
 791         /* Outer loop uses 7 flops */
 792     }
 793
 794     /* Increment number of outer iterations */
 795     outeriter        += nri;
 796
 797     /* Update outer/inner flops */
 798
 799     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*65);
 800 }