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, 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_ElecEwSw_VdwLJSw_GeomP1P1_VF_avx_128_fma_single
  54  * Electrostatics interaction: Ewald
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecEwSw_VdwLJSw_GeomP1P1_VF_avx_128_fma_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     __m128i          ewitab;
  99     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 100     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 101     real             *ewtab;
 102     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 103     real             rswitch_scalar,d_scalar;
 104     __m128           dummy_mask,cutoff_mask;
 105     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 106     __m128           one     = _mm_set1_ps(1.0);
 107     __m128           two     = _mm_set1_ps(2.0);
 108     x                = xx[0];
 109     f                = ff[0];
 110
 111     nri              = nlist->nri;
 112     iinr             = nlist->iinr;
 113     jindex           = nlist->jindex;
 114     jjnr             = nlist->jjnr;
 115     shiftidx         = nlist->shift;
 116     gid              = nlist->gid;
 117     shiftvec         = fr->shift_vec[0];
 118     fshift           = fr->fshift[0];
 119     facel            = _mm_set1_ps(fr->epsfac);
 120     charge           = mdatoms->chargeA;
 121     nvdwtype         = fr->ntype;
 122     vdwparam         = fr->nbfp;
 123     vdwtype          = mdatoms->typeA;
 124
 125     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 126     beta             = _mm_set1_ps(fr->ic->ewaldcoeff);
 127     beta2            = _mm_mul_ps(beta,beta);
 128     beta3            = _mm_mul_ps(beta,beta2);
 129     ewtab            = fr->ic->tabq_coul_FDV0;
 130     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 131     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 132
 133     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 134     rcutoff_scalar   = fr->rcoulomb;
 135     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 136     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 137
 138     rswitch_scalar   = fr->rcoulomb_switch;
 139     rswitch          = _mm_set1_ps(rswitch_scalar);
 140     /* Setup switch parameters */
 141     d_scalar         = rcutoff_scalar-rswitch_scalar;
 142     d                = _mm_set1_ps(d_scalar);
 143     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 144     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 145     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 146     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 147     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 148     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 149
 150     /* Avoid stupid compiler warnings */
 151     jnrA = jnrB = jnrC = jnrD = 0;
 152     j_coord_offsetA = 0;
 153     j_coord_offsetB = 0;
 154     j_coord_offsetC = 0;
 155     j_coord_offsetD = 0;
 156
 157     outeriter        = 0;
 158     inneriter        = 0;
 159
 160     for(iidx=0;iidx<4*DIM;iidx++)
 161     {
 162         scratch[iidx] = 0.0;
 163     }
 164
 165     /* Start outer loop over neighborlists */
 166     for(iidx=0; iidx<nri; iidx++)
 167     {
 168         /* Load shift vector for this list */
 169         i_shift_offset   = DIM*shiftidx[iidx];
 170
 171         /* Load limits for loop over neighbors */
 172         j_index_start    = jindex[iidx];
 173         j_index_end      = jindex[iidx+1];
 174
 175         /* Get outer coordinate index */
 176         inr              = iinr[iidx];
 177         i_coord_offset   = DIM*inr;
 178
 179         /* Load i particle coords and add shift vector */
 180         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 181
 182         fix0             = _mm_setzero_ps();
 183         fiy0             = _mm_setzero_ps();
 184         fiz0             = _mm_setzero_ps();
 185
 186         /* Load parameters for i particles */
 187         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 188         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 189
 190         /* Reset potential sums */
 191         velecsum         = _mm_setzero_ps();
 192         vvdwsum          = _mm_setzero_ps();
 193
 194         /* Start inner kernel loop */
 195         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 196         {
 197
 198             /* Get j neighbor index, and coordinate index */
 199             jnrA             = jjnr[jidx];
 200             jnrB             = jjnr[jidx+1];
 201             jnrC             = jjnr[jidx+2];
 202             jnrD             = jjnr[jidx+3];
 203             j_coord_offsetA  = DIM*jnrA;
 204             j_coord_offsetB  = DIM*jnrB;
 205             j_coord_offsetC  = DIM*jnrC;
 206             j_coord_offsetD  = DIM*jnrD;
 207
 208             /* load j atom coordinates */
 209             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 210                                               x+j_coord_offsetC,x+j_coord_offsetD,
 211                                               &jx0,&jy0,&jz0);
 212
 213             /* Calculate displacement vector */
 214             dx00             = _mm_sub_ps(ix0,jx0);
 215             dy00             = _mm_sub_ps(iy0,jy0);
 216             dz00             = _mm_sub_ps(iz0,jz0);
 217
 218             /* Calculate squared distance and things based on it */
 219             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 220
 221             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 222
 223             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 224
 225             /* Load parameters for j particles */
 226             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 227                                                               charge+jnrC+0,charge+jnrD+0);
 228             vdwjidx0A        = 2*vdwtype[jnrA+0];
 229             vdwjidx0B        = 2*vdwtype[jnrB+0];
 230             vdwjidx0C        = 2*vdwtype[jnrC+0];
 231             vdwjidx0D        = 2*vdwtype[jnrD+0];
 232
 233             /**************************
 234              * CALCULATE INTERACTIONS *
 235              **************************/
 236
 237             if (gmx_mm_any_lt(rsq00,rcutoff2))
 238             {
 239
 240             r00              = _mm_mul_ps(rsq00,rinv00);
 241
 242             /* Compute parameters for interactions between i and j atoms */
 243             qq00             = _mm_mul_ps(iq0,jq0);
 244             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 245                                          vdwparam+vdwioffset0+vdwjidx0B,
 246                                          vdwparam+vdwioffset0+vdwjidx0C,
 247                                          vdwparam+vdwioffset0+vdwjidx0D,
 248                                          &c6_00,&c12_00);
 249
 250             /* EWALD ELECTROSTATICS */
 251
 252             /* Analytical PME correction */
 253             zeta2            = _mm_mul_ps(beta2,rsq00);
 254             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 255             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 256             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 257             felec            = _mm_mul_ps(qq00,felec);
 258             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 259             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 260             velec            = _mm_mul_ps(qq00,velec);
 261
 262             /* LENNARD-JONES DISPERSION/REPULSION */
 263
 264             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 265             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 266             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 267             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 268             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 269
 270             d                = _mm_sub_ps(r00,rswitch);
 271             d                = _mm_max_ps(d,_mm_setzero_ps());
 272             d2               = _mm_mul_ps(d,d);
 273             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 274
 275             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 276
 277             /* Evaluate switch function */
 278             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 279             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 280             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 281             velec            = _mm_mul_ps(velec,sw);
 282             vvdw             = _mm_mul_ps(vvdw,sw);
 283             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 284
 285             /* Update potential sum for this i atom from the interaction with this j atom. */
 286             velec            = _mm_and_ps(velec,cutoff_mask);
 287             velecsum         = _mm_add_ps(velecsum,velec);
 288             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 289             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 290
 291             fscal            = _mm_add_ps(felec,fvdw);
 292
 293             fscal            = _mm_and_ps(fscal,cutoff_mask);
 294
 295              /* Update vectorial force */
 296             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 297             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 298             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 299
 300             fjptrA             = f+j_coord_offsetA;
 301             fjptrB             = f+j_coord_offsetB;
 302             fjptrC             = f+j_coord_offsetC;
 303             fjptrD             = f+j_coord_offsetD;
 304             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 305                                                    _mm_mul_ps(dx00,fscal),
 306                                                    _mm_mul_ps(dy00,fscal),
 307                                                    _mm_mul_ps(dz00,fscal));
 308
 309             }
 310
 311             /* Inner loop uses 71 flops */
 312         }
 313
 314         if(jidx<j_index_end)
 315         {
 316
 317             /* Get j neighbor index, and coordinate index */
 318             jnrlistA         = jjnr[jidx];
 319             jnrlistB         = jjnr[jidx+1];
 320             jnrlistC         = jjnr[jidx+2];
 321             jnrlistD         = jjnr[jidx+3];
 322             /* Sign of each element will be negative for non-real atoms.
 323              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 324              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 325              */
 326             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 327             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 328             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 329             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 330             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 331             j_coord_offsetA  = DIM*jnrA;
 332             j_coord_offsetB  = DIM*jnrB;
 333             j_coord_offsetC  = DIM*jnrC;
 334             j_coord_offsetD  = DIM*jnrD;
 335
 336             /* load j atom coordinates */
 337             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 338                                               x+j_coord_offsetC,x+j_coord_offsetD,
 339                                               &jx0,&jy0,&jz0);
 340
 341             /* Calculate displacement vector */
 342             dx00             = _mm_sub_ps(ix0,jx0);
 343             dy00             = _mm_sub_ps(iy0,jy0);
 344             dz00             = _mm_sub_ps(iz0,jz0);
 345
 346             /* Calculate squared distance and things based on it */
 347             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 348
 349             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 350
 351             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 352
 353             /* Load parameters for j particles */
 354             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 355                                                               charge+jnrC+0,charge+jnrD+0);
 356             vdwjidx0A        = 2*vdwtype[jnrA+0];
 357             vdwjidx0B        = 2*vdwtype[jnrB+0];
 358             vdwjidx0C        = 2*vdwtype[jnrC+0];
 359             vdwjidx0D        = 2*vdwtype[jnrD+0];
 360
 361             /**************************
 362              * CALCULATE INTERACTIONS *
 363              **************************/
 364
 365             if (gmx_mm_any_lt(rsq00,rcutoff2))
 366             {
 367
 368             r00              = _mm_mul_ps(rsq00,rinv00);
 369             r00              = _mm_andnot_ps(dummy_mask,r00);
 370
 371             /* Compute parameters for interactions between i and j atoms */
 372             qq00             = _mm_mul_ps(iq0,jq0);
 373             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 374                                          vdwparam+vdwioffset0+vdwjidx0B,
 375                                          vdwparam+vdwioffset0+vdwjidx0C,
 376                                          vdwparam+vdwioffset0+vdwjidx0D,
 377                                          &c6_00,&c12_00);
 378
 379             /* EWALD ELECTROSTATICS */
 380
 381             /* Analytical PME correction */
 382             zeta2            = _mm_mul_ps(beta2,rsq00);
 383             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 384             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 385             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 386             felec            = _mm_mul_ps(qq00,felec);
 387             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 388             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 389             velec            = _mm_mul_ps(qq00,velec);
 390
 391             /* LENNARD-JONES DISPERSION/REPULSION */
 392
 393             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 394             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 395             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 396             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 397             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 398
 399             d                = _mm_sub_ps(r00,rswitch);
 400             d                = _mm_max_ps(d,_mm_setzero_ps());
 401             d2               = _mm_mul_ps(d,d);
 402             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 403
 404             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 405
 406             /* Evaluate switch function */
 407             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 408             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 409             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 410             velec            = _mm_mul_ps(velec,sw);
 411             vvdw             = _mm_mul_ps(vvdw,sw);
 412             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 413
 414             /* Update potential sum for this i atom from the interaction with this j atom. */
 415             velec            = _mm_and_ps(velec,cutoff_mask);
 416             velec            = _mm_andnot_ps(dummy_mask,velec);
 417             velecsum         = _mm_add_ps(velecsum,velec);
 418             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 419             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 420             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 421
 422             fscal            = _mm_add_ps(felec,fvdw);
 423
 424             fscal            = _mm_and_ps(fscal,cutoff_mask);
 425
 426             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 427
 428              /* Update vectorial force */
 429             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 430             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 431             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 432
 433             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 434             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 435             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 436             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 437             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 438                                                    _mm_mul_ps(dx00,fscal),
 439                                                    _mm_mul_ps(dy00,fscal),
 440                                                    _mm_mul_ps(dz00,fscal));
 441
 442             }
 443
 444             /* Inner loop uses 72 flops */
 445         }
 446
 447         /* End of innermost loop */
 448
 449         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 450                                               f+i_coord_offset,fshift+i_shift_offset);
 451
 452         ggid                        = gid[iidx];
 453         /* Update potential energies */
 454         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 455         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 456
 457         /* Increment number of inner iterations */
 458         inneriter                  += j_index_end - j_index_start;
 459
 460         /* Outer loop uses 9 flops */
 461     }
 462
 463     /* Increment number of outer iterations */
 464     outeriter        += nri;
 465
 466     /* Update outer/inner flops */
 467
 468     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*72);
 469 }
 470 /*
 471  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSw_VdwLJSw_GeomP1P1_F_avx_128_fma_single
 472  * Electrostatics interaction: Ewald
 473  * VdW interaction:            LennardJones
 474  * Geometry:                   Particle-Particle
 475  * Calculate force/pot:        Force
 476  */
 477 void
 478 nb_kernel_ElecEwSw_VdwLJSw_GeomP1P1_F_avx_128_fma_single
 479                     (t_nblist                    * gmx_restrict       nlist,
 480                      rvec                        * gmx_restrict          xx,
 481                      rvec                        * gmx_restrict          ff,
 482                      t_forcerec                  * gmx_restrict          fr,
 483                      t_mdatoms                   * gmx_restrict     mdatoms,
 484                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 485                      t_nrnb                      * gmx_restrict        nrnb)
 486 {
 487     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 488      * just 0 for non-waters.
 489      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 490      * jnr indices corresponding to data put in the four positions in the SIMD register.
 491      */
 492     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 493     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 494     int              jnrA,jnrB,jnrC,jnrD;
 495     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 496     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 497     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 498     real             rcutoff_scalar;
 499     real             *shiftvec,*fshift,*x,*f;
 500     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 501     real             scratch[4*DIM];
 502     __m128           fscal,rcutoff,rcutoff2,jidxall;
 503     int              vdwioffset0;
 504     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 505     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 506     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 507     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 508     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 509     real             *charge;
 510     int              nvdwtype;
 511     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 512     int              *vdwtype;
 513     real             *vdwparam;
 514     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 515     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 516     __m128i          ewitab;
 517     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 518     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 519     real             *ewtab;
 520     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 521     real             rswitch_scalar,d_scalar;
 522     __m128           dummy_mask,cutoff_mask;
 523     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 524     __m128           one     = _mm_set1_ps(1.0);
 525     __m128           two     = _mm_set1_ps(2.0);
 526     x                = xx[0];
 527     f                = ff[0];
 528
 529     nri              = nlist->nri;
 530     iinr             = nlist->iinr;
 531     jindex           = nlist->jindex;
 532     jjnr             = nlist->jjnr;
 533     shiftidx         = nlist->shift;
 534     gid              = nlist->gid;
 535     shiftvec         = fr->shift_vec[0];
 536     fshift           = fr->fshift[0];
 537     facel            = _mm_set1_ps(fr->epsfac);
 538     charge           = mdatoms->chargeA;
 539     nvdwtype         = fr->ntype;
 540     vdwparam         = fr->nbfp;
 541     vdwtype          = mdatoms->typeA;
 542
 543     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 544     beta             = _mm_set1_ps(fr->ic->ewaldcoeff);
 545     beta2            = _mm_mul_ps(beta,beta);
 546     beta3            = _mm_mul_ps(beta,beta2);
 547     ewtab            = fr->ic->tabq_coul_FDV0;
 548     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 549     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 550
 551     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 552     rcutoff_scalar   = fr->rcoulomb;
 553     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 554     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 555
 556     rswitch_scalar   = fr->rcoulomb_switch;
 557     rswitch          = _mm_set1_ps(rswitch_scalar);
 558     /* Setup switch parameters */
 559     d_scalar         = rcutoff_scalar-rswitch_scalar;
 560     d                = _mm_set1_ps(d_scalar);
 561     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 562     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 563     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 564     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 565     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 566     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 567
 568     /* Avoid stupid compiler warnings */
 569     jnrA = jnrB = jnrC = jnrD = 0;
 570     j_coord_offsetA = 0;
 571     j_coord_offsetB = 0;
 572     j_coord_offsetC = 0;
 573     j_coord_offsetD = 0;
 574
 575     outeriter        = 0;
 576     inneriter        = 0;
 577
 578     for(iidx=0;iidx<4*DIM;iidx++)
 579     {
 580         scratch[iidx] = 0.0;
 581     }
 582
 583     /* Start outer loop over neighborlists */
 584     for(iidx=0; iidx<nri; iidx++)
 585     {
 586         /* Load shift vector for this list */
 587         i_shift_offset   = DIM*shiftidx[iidx];
 588
 589         /* Load limits for loop over neighbors */
 590         j_index_start    = jindex[iidx];
 591         j_index_end      = jindex[iidx+1];
 592
 593         /* Get outer coordinate index */
 594         inr              = iinr[iidx];
 595         i_coord_offset   = DIM*inr;
 596
 597         /* Load i particle coords and add shift vector */
 598         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 599
 600         fix0             = _mm_setzero_ps();
 601         fiy0             = _mm_setzero_ps();
 602         fiz0             = _mm_setzero_ps();
 603
 604         /* Load parameters for i particles */
 605         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 606         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 607
 608         /* Start inner kernel loop */
 609         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 610         {
 611
 612             /* Get j neighbor index, and coordinate index */
 613             jnrA             = jjnr[jidx];
 614             jnrB             = jjnr[jidx+1];
 615             jnrC             = jjnr[jidx+2];
 616             jnrD             = jjnr[jidx+3];
 617             j_coord_offsetA  = DIM*jnrA;
 618             j_coord_offsetB  = DIM*jnrB;
 619             j_coord_offsetC  = DIM*jnrC;
 620             j_coord_offsetD  = DIM*jnrD;
 621
 622             /* load j atom coordinates */
 623             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 624                                               x+j_coord_offsetC,x+j_coord_offsetD,
 625                                               &jx0,&jy0,&jz0);
 626
 627             /* Calculate displacement vector */
 628             dx00             = _mm_sub_ps(ix0,jx0);
 629             dy00             = _mm_sub_ps(iy0,jy0);
 630             dz00             = _mm_sub_ps(iz0,jz0);
 631
 632             /* Calculate squared distance and things based on it */
 633             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 634
 635             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 636
 637             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 638
 639             /* Load parameters for j particles */
 640             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 641                                                               charge+jnrC+0,charge+jnrD+0);
 642             vdwjidx0A        = 2*vdwtype[jnrA+0];
 643             vdwjidx0B        = 2*vdwtype[jnrB+0];
 644             vdwjidx0C        = 2*vdwtype[jnrC+0];
 645             vdwjidx0D        = 2*vdwtype[jnrD+0];
 646
 647             /**************************
 648              * CALCULATE INTERACTIONS *
 649              **************************/
 650
 651             if (gmx_mm_any_lt(rsq00,rcutoff2))
 652             {
 653
 654             r00              = _mm_mul_ps(rsq00,rinv00);
 655
 656             /* Compute parameters for interactions between i and j atoms */
 657             qq00             = _mm_mul_ps(iq0,jq0);
 658             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 659                                          vdwparam+vdwioffset0+vdwjidx0B,
 660                                          vdwparam+vdwioffset0+vdwjidx0C,
 661                                          vdwparam+vdwioffset0+vdwjidx0D,
 662                                          &c6_00,&c12_00);
 663
 664             /* EWALD ELECTROSTATICS */
 665
 666             /* Analytical PME correction */
 667             zeta2            = _mm_mul_ps(beta2,rsq00);
 668             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 669             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 670             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 671             felec            = _mm_mul_ps(qq00,felec);
 672             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 673             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 674             velec            = _mm_mul_ps(qq00,velec);
 675
 676             /* LENNARD-JONES DISPERSION/REPULSION */
 677
 678             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 679             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 680             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 681             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 682             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 683
 684             d                = _mm_sub_ps(r00,rswitch);
 685             d                = _mm_max_ps(d,_mm_setzero_ps());
 686             d2               = _mm_mul_ps(d,d);
 687             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 688
 689             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 690
 691             /* Evaluate switch function */
 692             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 693             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 694             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 695             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 696
 697             fscal            = _mm_add_ps(felec,fvdw);
 698
 699             fscal            = _mm_and_ps(fscal,cutoff_mask);
 700
 701              /* Update vectorial force */
 702             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 703             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 704             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 705
 706             fjptrA             = f+j_coord_offsetA;
 707             fjptrB             = f+j_coord_offsetB;
 708             fjptrC             = f+j_coord_offsetC;
 709             fjptrD             = f+j_coord_offsetD;
 710             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 711                                                    _mm_mul_ps(dx00,fscal),
 712                                                    _mm_mul_ps(dy00,fscal),
 713                                                    _mm_mul_ps(dz00,fscal));
 714
 715             }
 716
 717             /* Inner loop uses 65 flops */
 718         }
 719
 720         if(jidx<j_index_end)
 721         {
 722
 723             /* Get j neighbor index, and coordinate index */
 724             jnrlistA         = jjnr[jidx];
 725             jnrlistB         = jjnr[jidx+1];
 726             jnrlistC         = jjnr[jidx+2];
 727             jnrlistD         = jjnr[jidx+3];
 728             /* Sign of each element will be negative for non-real atoms.
 729              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 730              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 731              */
 732             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 733             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 734             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 735             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 736             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 737             j_coord_offsetA  = DIM*jnrA;
 738             j_coord_offsetB  = DIM*jnrB;
 739             j_coord_offsetC  = DIM*jnrC;
 740             j_coord_offsetD  = DIM*jnrD;
 741
 742             /* load j atom coordinates */
 743             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 744                                               x+j_coord_offsetC,x+j_coord_offsetD,
 745                                               &jx0,&jy0,&jz0);
 746
 747             /* Calculate displacement vector */
 748             dx00             = _mm_sub_ps(ix0,jx0);
 749             dy00             = _mm_sub_ps(iy0,jy0);
 750             dz00             = _mm_sub_ps(iz0,jz0);
 751
 752             /* Calculate squared distance and things based on it */
 753             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 754
 755             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 756
 757             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 758
 759             /* Load parameters for j particles */
 760             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 761                                                               charge+jnrC+0,charge+jnrD+0);
 762             vdwjidx0A        = 2*vdwtype[jnrA+0];
 763             vdwjidx0B        = 2*vdwtype[jnrB+0];
 764             vdwjidx0C        = 2*vdwtype[jnrC+0];
 765             vdwjidx0D        = 2*vdwtype[jnrD+0];
 766
 767             /**************************
 768              * CALCULATE INTERACTIONS *
 769              **************************/
 770
 771             if (gmx_mm_any_lt(rsq00,rcutoff2))
 772             {
 773
 774             r00              = _mm_mul_ps(rsq00,rinv00);
 775             r00              = _mm_andnot_ps(dummy_mask,r00);
 776
 777             /* Compute parameters for interactions between i and j atoms */
 778             qq00             = _mm_mul_ps(iq0,jq0);
 779             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 780                                          vdwparam+vdwioffset0+vdwjidx0B,
 781                                          vdwparam+vdwioffset0+vdwjidx0C,
 782                                          vdwparam+vdwioffset0+vdwjidx0D,
 783                                          &c6_00,&c12_00);
 784
 785             /* EWALD ELECTROSTATICS */
 786
 787             /* Analytical PME correction */
 788             zeta2            = _mm_mul_ps(beta2,rsq00);
 789             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 790             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 791             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 792             felec            = _mm_mul_ps(qq00,felec);
 793             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 794             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 795             velec            = _mm_mul_ps(qq00,velec);
 796
 797             /* LENNARD-JONES DISPERSION/REPULSION */
 798
 799             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 800             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 801             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 802             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 803             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 804
 805             d                = _mm_sub_ps(r00,rswitch);
 806             d                = _mm_max_ps(d,_mm_setzero_ps());
 807             d2               = _mm_mul_ps(d,d);
 808             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 809
 810             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 811
 812             /* Evaluate switch function */
 813             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 814             felec            = _mm_msub_ps( felec,sw , _mm_mul_ps(rinv00,_mm_mul_ps(velec,dsw)) );
 815             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 816             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 817
 818             fscal            = _mm_add_ps(felec,fvdw);
 819
 820             fscal            = _mm_and_ps(fscal,cutoff_mask);
 821
 822             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 823
 824              /* Update vectorial force */
 825             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 826             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 827             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 828
 829             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 830             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 831             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 832             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 833             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 834                                                    _mm_mul_ps(dx00,fscal),
 835                                                    _mm_mul_ps(dy00,fscal),
 836                                                    _mm_mul_ps(dz00,fscal));
 837
 838             }
 839
 840             /* Inner loop uses 66 flops */
 841         }
 842
 843         /* End of innermost loop */
 844
 845         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 846                                               f+i_coord_offset,fshift+i_shift_offset);
 847
 848         /* Increment number of inner iterations */
 849         inneriter                  += j_index_end - j_index_start;
 850
 851         /* Outer loop uses 7 flops */
 852     }
 853
 854     /* Increment number of outer iterations */
 855     outeriter        += nri;
 856
 857     /* Update outer/inner flops */
 858
 859     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*66);
 860 }