src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEwSw_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_ElecEwSw_VdwLJSw_GeomP1P1_VF_avx_128_fma_single
  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_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     __m128i          ewitab;
  97     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  98     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
  99     real             *ewtab;
 100     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 101     real             rswitch_scalar,d_scalar;
 102     __m128           dummy_mask,cutoff_mask;
 103     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 104     __m128           one     = _mm_set1_ps(1.0);
 105     __m128           two     = _mm_set1_ps(2.0);
 106     x                = xx[0];
 107     f                = ff[0];
 108
 109     nri              = nlist->nri;
 110     iinr             = nlist->iinr;
 111     jindex           = nlist->jindex;
 112     jjnr             = nlist->jjnr;
 113     shiftidx         = nlist->shift;
 114     gid              = nlist->gid;
 115     shiftvec         = fr->shift_vec[0];
 116     fshift           = fr->fshift[0];
 117     facel            = _mm_set1_ps(fr->epsfac);
 118     charge           = mdatoms->chargeA;
 119     nvdwtype         = fr->ntype;
 120     vdwparam         = fr->nbfp;
 121     vdwtype          = mdatoms->typeA;
 122
 123     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 124     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 125     beta2            = _mm_mul_ps(beta,beta);
 126     beta3            = _mm_mul_ps(beta,beta2);
 127     ewtab            = fr->ic->tabq_coul_FDV0;
 128     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 129     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 130
 131     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 132     rcutoff_scalar   = fr->rcoulomb;
 133     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 134     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 135
 136     rswitch_scalar   = fr->rcoulomb_switch;
 137     rswitch          = _mm_set1_ps(rswitch_scalar);
 138     /* Setup switch parameters */
 139     d_scalar         = rcutoff_scalar-rswitch_scalar;
 140     d                = _mm_set1_ps(d_scalar);
 141     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 142     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 143     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 144     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 145     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 146     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 147
 148     /* Avoid stupid compiler warnings */
 149     jnrA = jnrB = jnrC = jnrD = 0;
 150     j_coord_offsetA = 0;
 151     j_coord_offsetB = 0;
 152     j_coord_offsetC = 0;
 153     j_coord_offsetD = 0;
 154
 155     outeriter        = 0;
 156     inneriter        = 0;
 157
 158     for(iidx=0;iidx<4*DIM;iidx++)
 159     {
 160         scratch[iidx] = 0.0;
 161     }
 162
 163     /* Start outer loop over neighborlists */
 164     for(iidx=0; iidx<nri; iidx++)
 165     {
 166         /* Load shift vector for this list */
 167         i_shift_offset   = DIM*shiftidx[iidx];
 168
 169         /* Load limits for loop over neighbors */
 170         j_index_start    = jindex[iidx];
 171         j_index_end      = jindex[iidx+1];
 172
 173         /* Get outer coordinate index */
 174         inr              = iinr[iidx];
 175         i_coord_offset   = DIM*inr;
 176
 177         /* Load i particle coords and add shift vector */
 178         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 179
 180         fix0             = _mm_setzero_ps();
 181         fiy0             = _mm_setzero_ps();
 182         fiz0             = _mm_setzero_ps();
 183
 184         /* Load parameters for i particles */
 185         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 186         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 187
 188         /* Reset potential sums */
 189         velecsum         = _mm_setzero_ps();
 190         vvdwsum          = _mm_setzero_ps();
 191
 192         /* Start inner kernel loop */
 193         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 194         {
 195
 196             /* Get j neighbor index, and coordinate index */
 197             jnrA             = jjnr[jidx];
 198             jnrB             = jjnr[jidx+1];
 199             jnrC             = jjnr[jidx+2];
 200             jnrD             = jjnr[jidx+3];
 201             j_coord_offsetA  = DIM*jnrA;
 202             j_coord_offsetB  = DIM*jnrB;
 203             j_coord_offsetC  = DIM*jnrC;
 204             j_coord_offsetD  = DIM*jnrD;
 205
 206             /* load j atom coordinates */
 207             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 208                                               x+j_coord_offsetC,x+j_coord_offsetD,
 209                                               &jx0,&jy0,&jz0);
 210
 211             /* Calculate displacement vector */
 212             dx00             = _mm_sub_ps(ix0,jx0);
 213             dy00             = _mm_sub_ps(iy0,jy0);
 214             dz00             = _mm_sub_ps(iz0,jz0);
 215
 216             /* Calculate squared distance and things based on it */
 217             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 218
 219             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 220
 221             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 222
 223             /* Load parameters for j particles */
 224             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 225                                                               charge+jnrC+0,charge+jnrD+0);
 226             vdwjidx0A        = 2*vdwtype[jnrA+0];
 227             vdwjidx0B        = 2*vdwtype[jnrB+0];
 228             vdwjidx0C        = 2*vdwtype[jnrC+0];
 229             vdwjidx0D        = 2*vdwtype[jnrD+0];
 230
 231             /**************************
 232              * CALCULATE INTERACTIONS *
 233              **************************/
 234
 235             if (gmx_mm_any_lt(rsq00,rcutoff2))
 236             {
 237
 238             r00              = _mm_mul_ps(rsq00,rinv00);
 239
 240             /* Compute parameters for interactions between i and j atoms */
 241             qq00             = _mm_mul_ps(iq0,jq0);
 242             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 243                                          vdwparam+vdwioffset0+vdwjidx0B,
 244                                          vdwparam+vdwioffset0+vdwjidx0C,
 245                                          vdwparam+vdwioffset0+vdwjidx0D,
 246                                          &c6_00,&c12_00);
 247
 248             /* EWALD ELECTROSTATICS */
 249
 250             /* Analytical PME correction */
 251             zeta2            = _mm_mul_ps(beta2,rsq00);
 252             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 253             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 254             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 255             felec            = _mm_mul_ps(qq00,felec);
 256             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 257             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 258             velec            = _mm_mul_ps(qq00,velec);
 259
 260             /* LENNARD-JONES DISPERSION/REPULSION */
 261
 262             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 263             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 264             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 265             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 266             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 267
 268             d                = _mm_sub_ps(r00,rswitch);
 269             d                = _mm_max_ps(d,_mm_setzero_ps());
 270             d2               = _mm_mul_ps(d,d);
 271             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 272
 273             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 274
 275             /* Evaluate switch function */
 276             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 277             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 278             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 279             velec            = _mm_mul_ps(velec,sw);
 280             vvdw             = _mm_mul_ps(vvdw,sw);
 281             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 282
 283             /* Update potential sum for this i atom from the interaction with this j atom. */
 284             velec            = _mm_and_ps(velec,cutoff_mask);
 285             velecsum         = _mm_add_ps(velecsum,velec);
 286             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 287             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 288
 289             fscal            = _mm_add_ps(felec,fvdw);
 290
 291             fscal            = _mm_and_ps(fscal,cutoff_mask);
 292
 293              /* Update vectorial force */
 294             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 295             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 296             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 297
 298             fjptrA             = f+j_coord_offsetA;
 299             fjptrB             = f+j_coord_offsetB;
 300             fjptrC             = f+j_coord_offsetC;
 301             fjptrD             = f+j_coord_offsetD;
 302             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 303                                                    _mm_mul_ps(dx00,fscal),
 304                                                    _mm_mul_ps(dy00,fscal),
 305                                                    _mm_mul_ps(dz00,fscal));
 306
 307             }
 308
 309             /* Inner loop uses 71 flops */
 310         }
 311
 312         if(jidx<j_index_end)
 313         {
 314
 315             /* Get j neighbor index, and coordinate index */
 316             jnrlistA         = jjnr[jidx];
 317             jnrlistB         = jjnr[jidx+1];
 318             jnrlistC         = jjnr[jidx+2];
 319             jnrlistD         = jjnr[jidx+3];
 320             /* Sign of each element will be negative for non-real atoms.
 321              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 322              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 323              */
 324             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 325             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 326             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 327             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 328             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 329             j_coord_offsetA  = DIM*jnrA;
 330             j_coord_offsetB  = DIM*jnrB;
 331             j_coord_offsetC  = DIM*jnrC;
 332             j_coord_offsetD  = DIM*jnrD;
 333
 334             /* load j atom coordinates */
 335             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 336                                               x+j_coord_offsetC,x+j_coord_offsetD,
 337                                               &jx0,&jy0,&jz0);
 338
 339             /* Calculate displacement vector */
 340             dx00             = _mm_sub_ps(ix0,jx0);
 341             dy00             = _mm_sub_ps(iy0,jy0);
 342             dz00             = _mm_sub_ps(iz0,jz0);
 343
 344             /* Calculate squared distance and things based on it */
 345             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 346
 347             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 348
 349             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 350
 351             /* Load parameters for j particles */
 352             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 353                                                               charge+jnrC+0,charge+jnrD+0);
 354             vdwjidx0A        = 2*vdwtype[jnrA+0];
 355             vdwjidx0B        = 2*vdwtype[jnrB+0];
 356             vdwjidx0C        = 2*vdwtype[jnrC+0];
 357             vdwjidx0D        = 2*vdwtype[jnrD+0];
 358
 359             /**************************
 360              * CALCULATE INTERACTIONS *
 361              **************************/
 362
 363             if (gmx_mm_any_lt(rsq00,rcutoff2))
 364             {
 365
 366             r00              = _mm_mul_ps(rsq00,rinv00);
 367             r00              = _mm_andnot_ps(dummy_mask,r00);
 368
 369             /* Compute parameters for interactions between i and j atoms */
 370             qq00             = _mm_mul_ps(iq0,jq0);
 371             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 372                                          vdwparam+vdwioffset0+vdwjidx0B,
 373                                          vdwparam+vdwioffset0+vdwjidx0C,
 374                                          vdwparam+vdwioffset0+vdwjidx0D,
 375                                          &c6_00,&c12_00);
 376
 377             /* EWALD ELECTROSTATICS */
 378
 379             /* Analytical PME correction */
 380             zeta2            = _mm_mul_ps(beta2,rsq00);
 381             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 382             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 383             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 384             felec            = _mm_mul_ps(qq00,felec);
 385             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 386             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 387             velec            = _mm_mul_ps(qq00,velec);
 388
 389             /* LENNARD-JONES DISPERSION/REPULSION */
 390
 391             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 392             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 393             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 394             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 395             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 396
 397             d                = _mm_sub_ps(r00,rswitch);
 398             d                = _mm_max_ps(d,_mm_setzero_ps());
 399             d2               = _mm_mul_ps(d,d);
 400             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 401
 402             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 403
 404             /* Evaluate switch function */
 405             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 406             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 407             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 408             velec            = _mm_mul_ps(velec,sw);
 409             vvdw             = _mm_mul_ps(vvdw,sw);
 410             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 411
 412             /* Update potential sum for this i atom from the interaction with this j atom. */
 413             velec            = _mm_and_ps(velec,cutoff_mask);
 414             velec            = _mm_andnot_ps(dummy_mask,velec);
 415             velecsum         = _mm_add_ps(velecsum,velec);
 416             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 417             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 418             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 419
 420             fscal            = _mm_add_ps(felec,fvdw);
 421
 422             fscal            = _mm_and_ps(fscal,cutoff_mask);
 423
 424             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 425
 426              /* Update vectorial force */
 427             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 428             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 429             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 430
 431             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 432             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 433             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 434             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 435             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 436                                                    _mm_mul_ps(dx00,fscal),
 437                                                    _mm_mul_ps(dy00,fscal),
 438                                                    _mm_mul_ps(dz00,fscal));
 439
 440             }
 441
 442             /* Inner loop uses 72 flops */
 443         }
 444
 445         /* End of innermost loop */
 446
 447         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 448                                               f+i_coord_offset,fshift+i_shift_offset);
 449
 450         ggid                        = gid[iidx];
 451         /* Update potential energies */
 452         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 453         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 454
 455         /* Increment number of inner iterations */
 456         inneriter                  += j_index_end - j_index_start;
 457
 458         /* Outer loop uses 9 flops */
 459     }
 460
 461     /* Increment number of outer iterations */
 462     outeriter        += nri;
 463
 464     /* Update outer/inner flops */
 465
 466     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*72);
 467 }
 468 /*
 469  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwLJSw_GeomP1P1_F_avx_128_fma_single
 470  * Electrostatics interaction: Ewald
 471  * VdW interaction:            LennardJones
 472  * Geometry:                   Particle-Particle
 473  * Calculate force/pot:        Force
 474  */
 475 void
 476 nb_kernel_ElecEwSw_VdwLJSw_GeomP1P1_F_avx_128_fma_single
 477                     (t_nblist                    * gmx_restrict       nlist,
 478                      rvec                        * gmx_restrict          xx,
 479                      rvec                        * gmx_restrict          ff,
 480                      t_forcerec                  * gmx_restrict          fr,
 481                      t_mdatoms                   * gmx_restrict     mdatoms,
 482                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 483                      t_nrnb                      * gmx_restrict        nrnb)
 484 {
 485     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 486      * just 0 for non-waters.
 487      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 488      * jnr indices corresponding to data put in the four positions in the SIMD register.
 489      */
 490     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 491     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 492     int              jnrA,jnrB,jnrC,jnrD;
 493     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 494     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 495     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 496     real             rcutoff_scalar;
 497     real             *shiftvec,*fshift,*x,*f;
 498     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 499     real             scratch[4*DIM];
 500     __m128           fscal,rcutoff,rcutoff2,jidxall;
 501     int              vdwioffset0;
 502     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 503     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 504     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 505     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 506     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 507     real             *charge;
 508     int              nvdwtype;
 509     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 510     int              *vdwtype;
 511     real             *vdwparam;
 512     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 513     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 514     __m128i          ewitab;
 515     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 516     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 517     real             *ewtab;
 518     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 519     real             rswitch_scalar,d_scalar;
 520     __m128           dummy_mask,cutoff_mask;
 521     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 522     __m128           one     = _mm_set1_ps(1.0);
 523     __m128           two     = _mm_set1_ps(2.0);
 524     x                = xx[0];
 525     f                = ff[0];
 526
 527     nri              = nlist->nri;
 528     iinr             = nlist->iinr;
 529     jindex           = nlist->jindex;
 530     jjnr             = nlist->jjnr;
 531     shiftidx         = nlist->shift;
 532     gid              = nlist->gid;
 533     shiftvec         = fr->shift_vec[0];
 534     fshift           = fr->fshift[0];
 535     facel            = _mm_set1_ps(fr->epsfac);
 536     charge           = mdatoms->chargeA;
 537     nvdwtype         = fr->ntype;
 538     vdwparam         = fr->nbfp;
 539     vdwtype          = mdatoms->typeA;
 540
 541     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 542     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 543     beta2            = _mm_mul_ps(beta,beta);
 544     beta3            = _mm_mul_ps(beta,beta2);
 545     ewtab            = fr->ic->tabq_coul_FDV0;
 546     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 547     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 548
 549     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 550     rcutoff_scalar   = fr->rcoulomb;
 551     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 552     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 553
 554     rswitch_scalar   = fr->rcoulomb_switch;
 555     rswitch          = _mm_set1_ps(rswitch_scalar);
 556     /* Setup switch parameters */
 557     d_scalar         = rcutoff_scalar-rswitch_scalar;
 558     d                = _mm_set1_ps(d_scalar);
 559     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 560     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 561     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 562     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 563     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 564     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 565
 566     /* Avoid stupid compiler warnings */
 567     jnrA = jnrB = jnrC = jnrD = 0;
 568     j_coord_offsetA = 0;
 569     j_coord_offsetB = 0;
 570     j_coord_offsetC = 0;
 571     j_coord_offsetD = 0;
 572
 573     outeriter        = 0;
 574     inneriter        = 0;
 575
 576     for(iidx=0;iidx<4*DIM;iidx++)
 577     {
 578         scratch[iidx] = 0.0;
 579     }
 580
 581     /* Start outer loop over neighborlists */
 582     for(iidx=0; iidx<nri; iidx++)
 583     {
 584         /* Load shift vector for this list */
 585         i_shift_offset   = DIM*shiftidx[iidx];
 586
 587         /* Load limits for loop over neighbors */
 588         j_index_start    = jindex[iidx];
 589         j_index_end      = jindex[iidx+1];
 590
 591         /* Get outer coordinate index */
 592         inr              = iinr[iidx];
 593         i_coord_offset   = DIM*inr;
 594
 595         /* Load i particle coords and add shift vector */
 596         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 597
 598         fix0             = _mm_setzero_ps();
 599         fiy0             = _mm_setzero_ps();
 600         fiz0             = _mm_setzero_ps();
 601
 602         /* Load parameters for i particles */
 603         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 604         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 605
 606         /* Start inner kernel loop */
 607         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 608         {
 609
 610             /* Get j neighbor index, and coordinate index */
 611             jnrA             = jjnr[jidx];
 612             jnrB             = jjnr[jidx+1];
 613             jnrC             = jjnr[jidx+2];
 614             jnrD             = jjnr[jidx+3];
 615             j_coord_offsetA  = DIM*jnrA;
 616             j_coord_offsetB  = DIM*jnrB;
 617             j_coord_offsetC  = DIM*jnrC;
 618             j_coord_offsetD  = DIM*jnrD;
 619
 620             /* load j atom coordinates */
 621             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 622                                               x+j_coord_offsetC,x+j_coord_offsetD,
 623                                               &jx0,&jy0,&jz0);
 624
 625             /* Calculate displacement vector */
 626             dx00             = _mm_sub_ps(ix0,jx0);
 627             dy00             = _mm_sub_ps(iy0,jy0);
 628             dz00             = _mm_sub_ps(iz0,jz0);
 629
 630             /* Calculate squared distance and things based on it */
 631             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 632
 633             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 634
 635             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 636
 637             /* Load parameters for j particles */
 638             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 639                                                               charge+jnrC+0,charge+jnrD+0);
 640             vdwjidx0A        = 2*vdwtype[jnrA+0];
 641             vdwjidx0B        = 2*vdwtype[jnrB+0];
 642             vdwjidx0C        = 2*vdwtype[jnrC+0];
 643             vdwjidx0D        = 2*vdwtype[jnrD+0];
 644
 645             /**************************
 646              * CALCULATE INTERACTIONS *
 647              **************************/
 648
 649             if (gmx_mm_any_lt(rsq00,rcutoff2))
 650             {
 651
 652             r00              = _mm_mul_ps(rsq00,rinv00);
 653
 654             /* Compute parameters for interactions between i and j atoms */
 655             qq00             = _mm_mul_ps(iq0,jq0);
 656             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 657                                          vdwparam+vdwioffset0+vdwjidx0B,
 658                                          vdwparam+vdwioffset0+vdwjidx0C,
 659                                          vdwparam+vdwioffset0+vdwjidx0D,
 660                                          &c6_00,&c12_00);
 661
 662             /* EWALD ELECTROSTATICS */
 663
 664             /* Analytical PME correction */
 665             zeta2            = _mm_mul_ps(beta2,rsq00);
 666             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 667             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 668             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 669             felec            = _mm_mul_ps(qq00,felec);
 670             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 671             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 672             velec            = _mm_mul_ps(qq00,velec);
 673
 674             /* LENNARD-JONES DISPERSION/REPULSION */
 675
 676             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 677             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 678             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 679             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 680             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 681
 682             d                = _mm_sub_ps(r00,rswitch);
 683             d                = _mm_max_ps(d,_mm_setzero_ps());
 684             d2               = _mm_mul_ps(d,d);
 685             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 686
 687             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 688
 689             /* Evaluate switch function */
 690             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 691             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 692             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 693             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 694
 695             fscal            = _mm_add_ps(felec,fvdw);
 696
 697             fscal            = _mm_and_ps(fscal,cutoff_mask);
 698
 699              /* Update vectorial force */
 700             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 701             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 702             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 703
 704             fjptrA             = f+j_coord_offsetA;
 705             fjptrB             = f+j_coord_offsetB;
 706             fjptrC             = f+j_coord_offsetC;
 707             fjptrD             = f+j_coord_offsetD;
 708             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 709                                                    _mm_mul_ps(dx00,fscal),
 710                                                    _mm_mul_ps(dy00,fscal),
 711                                                    _mm_mul_ps(dz00,fscal));
 712
 713             }
 714
 715             /* Inner loop uses 65 flops */
 716         }
 717
 718         if(jidx<j_index_end)
 719         {
 720
 721             /* Get j neighbor index, and coordinate index */
 722             jnrlistA         = jjnr[jidx];
 723             jnrlistB         = jjnr[jidx+1];
 724             jnrlistC         = jjnr[jidx+2];
 725             jnrlistD         = jjnr[jidx+3];
 726             /* Sign of each element will be negative for non-real atoms.
 727              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 728              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 729              */
 730             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 731             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 732             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 733             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 734             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 735             j_coord_offsetA  = DIM*jnrA;
 736             j_coord_offsetB  = DIM*jnrB;
 737             j_coord_offsetC  = DIM*jnrC;
 738             j_coord_offsetD  = DIM*jnrD;
 739
 740             /* load j atom coordinates */
 741             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 742                                               x+j_coord_offsetC,x+j_coord_offsetD,
 743                                               &jx0,&jy0,&jz0);
 744
 745             /* Calculate displacement vector */
 746             dx00             = _mm_sub_ps(ix0,jx0);
 747             dy00             = _mm_sub_ps(iy0,jy0);
 748             dz00             = _mm_sub_ps(iz0,jz0);
 749
 750             /* Calculate squared distance and things based on it */
 751             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 752
 753             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 754
 755             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 756
 757             /* Load parameters for j particles */
 758             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 759                                                               charge+jnrC+0,charge+jnrD+0);
 760             vdwjidx0A        = 2*vdwtype[jnrA+0];
 761             vdwjidx0B        = 2*vdwtype[jnrB+0];
 762             vdwjidx0C        = 2*vdwtype[jnrC+0];
 763             vdwjidx0D        = 2*vdwtype[jnrD+0];
 764
 765             /**************************
 766              * CALCULATE INTERACTIONS *
 767              **************************/
 768
 769             if (gmx_mm_any_lt(rsq00,rcutoff2))
 770             {
 771
 772             r00              = _mm_mul_ps(rsq00,rinv00);
 773             r00              = _mm_andnot_ps(dummy_mask,r00);
 774
 775             /* Compute parameters for interactions between i and j atoms */
 776             qq00             = _mm_mul_ps(iq0,jq0);
 777             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 778                                          vdwparam+vdwioffset0+vdwjidx0B,
 779                                          vdwparam+vdwioffset0+vdwjidx0C,
 780                                          vdwparam+vdwioffset0+vdwjidx0D,
 781                                          &c6_00,&c12_00);
 782
 783             /* EWALD ELECTROSTATICS */
 784
 785             /* Analytical PME correction */
 786             zeta2            = _mm_mul_ps(beta2,rsq00);
 787             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 788             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 789             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 790             felec            = _mm_mul_ps(qq00,felec);
 791             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 792             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 793             velec            = _mm_mul_ps(qq00,velec);
 794
 795             /* LENNARD-JONES DISPERSION/REPULSION */
 796
 797             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 798             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 799             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 800             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 801             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 802
 803             d                = _mm_sub_ps(r00,rswitch);
 804             d                = _mm_max_ps(d,_mm_setzero_ps());
 805             d2               = _mm_mul_ps(d,d);
 806             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 807
 808             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 809
 810             /* Evaluate switch function */
 811             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 812             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 813             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 814             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 815
 816             fscal            = _mm_add_ps(felec,fvdw);
 817
 818             fscal            = _mm_and_ps(fscal,cutoff_mask);
 819
 820             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 821
 822              /* Update vectorial force */
 823             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 824             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 825             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 826
 827             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 828             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 829             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 830             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 831             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 832                                                    _mm_mul_ps(dx00,fscal),
 833                                                    _mm_mul_ps(dy00,fscal),
 834                                                    _mm_mul_ps(dz00,fscal));
 835
 836             }
 837
 838             /* Inner loop uses 66 flops */
 839         }
 840
 841         /* End of innermost loop */
 842
 843         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 844                                               f+i_coord_offset,fshift+i_shift_offset);
 845
 846         /* Increment number of inner iterations */
 847         inneriter                  += j_index_end - j_index_start;
 848
 849         /* Outer loop uses 7 flops */
 850     }
 851
 852     /* Increment number of outer iterations */
 853     outeriter        += nri;
 854
 855     /* Update outer/inner flops */
 856
 857     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*66);
 858 }