src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEwSh_VdwLJEwSh_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 "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "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_ElecEwSh_VdwLJEwSh_GeomP1P1_VF_avx_128_fma_single
  52  * Electrostatics interaction: Ewald
  53  * VdW interaction:            LJEwald
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecEwSh_VdwLJEwSh_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           c6grid_00;
  97     real             *vdwgridparam;
  98     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
  99     __m128           one_half = _mm_set1_ps(0.5);
 100     __m128           minus_one = _mm_set1_ps(-1.0);
 101     __m128i          ewitab;
 102     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 103     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 104     real             *ewtab;
 105     __m128           dummy_mask,cutoff_mask;
 106     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 107     __m128           one     = _mm_set1_ps(1.0);
 108     __m128           two     = _mm_set1_ps(2.0);
 109     x                = xx[0];
 110     f                = ff[0];
 111
 112     nri              = nlist->nri;
 113     iinr             = nlist->iinr;
 114     jindex           = nlist->jindex;
 115     jjnr             = nlist->jjnr;
 116     shiftidx         = nlist->shift;
 117     gid              = nlist->gid;
 118     shiftvec         = fr->shift_vec[0];
 119     fshift           = fr->fshift[0];
 120     facel            = _mm_set1_ps(fr->epsfac);
 121     charge           = mdatoms->chargeA;
 122     nvdwtype         = fr->ntype;
 123     vdwparam         = fr->nbfp;
 124     vdwtype          = mdatoms->typeA;
 125     vdwgridparam     = fr->ljpme_c6grid;
 126     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 127     ewclj            = _mm_set1_ps(fr->ewaldcoeff_lj);
 128     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 129
 130     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 131     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 132     beta2            = _mm_mul_ps(beta,beta);
 133     beta3            = _mm_mul_ps(beta,beta2);
 134     ewtab            = fr->ic->tabq_coul_FDV0;
 135     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 136     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 137
 138     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 139     rcutoff_scalar   = fr->rcoulomb;
 140     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 141     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 142
 143     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 144     rvdw             = _mm_set1_ps(fr->rvdw);
 145
 146     /* Avoid stupid compiler warnings */
 147     jnrA = jnrB = jnrC = jnrD = 0;
 148     j_coord_offsetA = 0;
 149     j_coord_offsetB = 0;
 150     j_coord_offsetC = 0;
 151     j_coord_offsetD = 0;
 152
 153     outeriter        = 0;
 154     inneriter        = 0;
 155
 156     for(iidx=0;iidx<4*DIM;iidx++)
 157     {
 158         scratch[iidx] = 0.0;
 159     }
 160
 161     /* Start outer loop over neighborlists */
 162     for(iidx=0; iidx<nri; iidx++)
 163     {
 164         /* Load shift vector for this list */
 165         i_shift_offset   = DIM*shiftidx[iidx];
 166
 167         /* Load limits for loop over neighbors */
 168         j_index_start    = jindex[iidx];
 169         j_index_end      = jindex[iidx+1];
 170
 171         /* Get outer coordinate index */
 172         inr              = iinr[iidx];
 173         i_coord_offset   = DIM*inr;
 174
 175         /* Load i particle coords and add shift vector */
 176         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 177
 178         fix0             = _mm_setzero_ps();
 179         fiy0             = _mm_setzero_ps();
 180         fiz0             = _mm_setzero_ps();
 181
 182         /* Load parameters for i particles */
 183         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 184         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 185
 186         /* Reset potential sums */
 187         velecsum         = _mm_setzero_ps();
 188         vvdwsum          = _mm_setzero_ps();
 189
 190         /* Start inner kernel loop */
 191         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 192         {
 193
 194             /* Get j neighbor index, and coordinate index */
 195             jnrA             = jjnr[jidx];
 196             jnrB             = jjnr[jidx+1];
 197             jnrC             = jjnr[jidx+2];
 198             jnrD             = jjnr[jidx+3];
 199             j_coord_offsetA  = DIM*jnrA;
 200             j_coord_offsetB  = DIM*jnrB;
 201             j_coord_offsetC  = DIM*jnrC;
 202             j_coord_offsetD  = DIM*jnrD;
 203
 204             /* load j atom coordinates */
 205             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 206                                               x+j_coord_offsetC,x+j_coord_offsetD,
 207                                               &jx0,&jy0,&jz0);
 208
 209             /* Calculate displacement vector */
 210             dx00             = _mm_sub_ps(ix0,jx0);
 211             dy00             = _mm_sub_ps(iy0,jy0);
 212             dz00             = _mm_sub_ps(iz0,jz0);
 213
 214             /* Calculate squared distance and things based on it */
 215             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 216
 217             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 218
 219             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 220
 221             /* Load parameters for j particles */
 222             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 223                                                               charge+jnrC+0,charge+jnrD+0);
 224             vdwjidx0A        = 2*vdwtype[jnrA+0];
 225             vdwjidx0B        = 2*vdwtype[jnrB+0];
 226             vdwjidx0C        = 2*vdwtype[jnrC+0];
 227             vdwjidx0D        = 2*vdwtype[jnrD+0];
 228
 229             /**************************
 230              * CALCULATE INTERACTIONS *
 231              **************************/
 232
 233             if (gmx_mm_any_lt(rsq00,rcutoff2))
 234             {
 235
 236             r00              = _mm_mul_ps(rsq00,rinv00);
 237
 238             /* Compute parameters for interactions between i and j atoms */
 239             qq00             = _mm_mul_ps(iq0,jq0);
 240             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 241                                          vdwparam+vdwioffset0+vdwjidx0B,
 242                                          vdwparam+vdwioffset0+vdwjidx0C,
 243                                          vdwparam+vdwioffset0+vdwjidx0D,
 244                                          &c6_00,&c12_00);
 245
 246             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 247                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 248                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 249                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 250
 251             /* EWALD ELECTROSTATICS */
 252
 253             /* Analytical PME correction */
 254             zeta2            = _mm_mul_ps(beta2,rsq00);
 255             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 256             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 257             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 258             felec            = _mm_mul_ps(qq00,felec);
 259             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 260             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv00,sh_ewald));
 261             velec            = _mm_mul_ps(qq00,velec);
 262
 263             /* Analytical LJ-PME */
 264             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 265             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 266             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 267             exponent         = gmx_simd_exp_r(ewcljrsq);
 268             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 269             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 270             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 271             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_00,_mm_sub_ps(one,poly),c6_00),rinvsix);
 272             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 273             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 274                                           _mm_mul_ps(_mm_sub_ps(vvdw6,_mm_macc_ps(c6grid_00,sh_lj_ewald,_mm_mul_ps(c6_00,sh_vdw_invrcut6))),one_sixth));
 275             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 276             fvdw             = _mm_mul_ps(_mm_add_ps(vvdw12,_mm_msub_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6),vvdw6)),rinvsq00);
 277
 278             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 279
 280             /* Update potential sum for this i atom from the interaction with this j atom. */
 281             velec            = _mm_and_ps(velec,cutoff_mask);
 282             velecsum         = _mm_add_ps(velecsum,velec);
 283             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 284             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 285
 286             fscal            = _mm_add_ps(felec,fvdw);
 287
 288             fscal            = _mm_and_ps(fscal,cutoff_mask);
 289
 290              /* Update vectorial force */
 291             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 292             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 293             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 294
 295             fjptrA             = f+j_coord_offsetA;
 296             fjptrB             = f+j_coord_offsetB;
 297             fjptrC             = f+j_coord_offsetC;
 298             fjptrD             = f+j_coord_offsetD;
 299             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 300                                                    _mm_mul_ps(dx00,fscal),
 301                                                    _mm_mul_ps(dy00,fscal),
 302                                                    _mm_mul_ps(dz00,fscal));
 303
 304             }
 305
 306             /* Inner loop uses 63 flops */
 307         }
 308
 309         if(jidx<j_index_end)
 310         {
 311
 312             /* Get j neighbor index, and coordinate index */
 313             jnrlistA         = jjnr[jidx];
 314             jnrlistB         = jjnr[jidx+1];
 315             jnrlistC         = jjnr[jidx+2];
 316             jnrlistD         = jjnr[jidx+3];
 317             /* Sign of each element will be negative for non-real atoms.
 318              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 319              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 320              */
 321             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 322             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 323             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 324             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 325             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 326             j_coord_offsetA  = DIM*jnrA;
 327             j_coord_offsetB  = DIM*jnrB;
 328             j_coord_offsetC  = DIM*jnrC;
 329             j_coord_offsetD  = DIM*jnrD;
 330
 331             /* load j atom coordinates */
 332             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 333                                               x+j_coord_offsetC,x+j_coord_offsetD,
 334                                               &jx0,&jy0,&jz0);
 335
 336             /* Calculate displacement vector */
 337             dx00             = _mm_sub_ps(ix0,jx0);
 338             dy00             = _mm_sub_ps(iy0,jy0);
 339             dz00             = _mm_sub_ps(iz0,jz0);
 340
 341             /* Calculate squared distance and things based on it */
 342             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 343
 344             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 345
 346             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 347
 348             /* Load parameters for j particles */
 349             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 350                                                               charge+jnrC+0,charge+jnrD+0);
 351             vdwjidx0A        = 2*vdwtype[jnrA+0];
 352             vdwjidx0B        = 2*vdwtype[jnrB+0];
 353             vdwjidx0C        = 2*vdwtype[jnrC+0];
 354             vdwjidx0D        = 2*vdwtype[jnrD+0];
 355
 356             /**************************
 357              * CALCULATE INTERACTIONS *
 358              **************************/
 359
 360             if (gmx_mm_any_lt(rsq00,rcutoff2))
 361             {
 362
 363             r00              = _mm_mul_ps(rsq00,rinv00);
 364             r00              = _mm_andnot_ps(dummy_mask,r00);
 365
 366             /* Compute parameters for interactions between i and j atoms */
 367             qq00             = _mm_mul_ps(iq0,jq0);
 368             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 369                                          vdwparam+vdwioffset0+vdwjidx0B,
 370                                          vdwparam+vdwioffset0+vdwjidx0C,
 371                                          vdwparam+vdwioffset0+vdwjidx0D,
 372                                          &c6_00,&c12_00);
 373
 374             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 375                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 376                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 377                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 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,_mm_sub_ps(rinv00,sh_ewald));
 389             velec            = _mm_mul_ps(qq00,velec);
 390
 391             /* Analytical LJ-PME */
 392             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 393             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 394             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 395             exponent         = gmx_simd_exp_r(ewcljrsq);
 396             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 397             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 398             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 399             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_00,_mm_sub_ps(one,poly),c6_00),rinvsix);
 400             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 401             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 402                                           _mm_mul_ps(_mm_sub_ps(vvdw6,_mm_macc_ps(c6grid_00,sh_lj_ewald,_mm_mul_ps(c6_00,sh_vdw_invrcut6))),one_sixth));
 403             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 404             fvdw             = _mm_mul_ps(_mm_add_ps(vvdw12,_mm_msub_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6),vvdw6)),rinvsq00);
 405
 406             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 407
 408             /* Update potential sum for this i atom from the interaction with this j atom. */
 409             velec            = _mm_and_ps(velec,cutoff_mask);
 410             velec            = _mm_andnot_ps(dummy_mask,velec);
 411             velecsum         = _mm_add_ps(velecsum,velec);
 412             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 413             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 414             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 415
 416             fscal            = _mm_add_ps(felec,fvdw);
 417
 418             fscal            = _mm_and_ps(fscal,cutoff_mask);
 419
 420             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 421
 422              /* Update vectorial force */
 423             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 424             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 425             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 426
 427             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 428             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 429             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 430             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 431             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 432                                                    _mm_mul_ps(dx00,fscal),
 433                                                    _mm_mul_ps(dy00,fscal),
 434                                                    _mm_mul_ps(dz00,fscal));
 435
 436             }
 437
 438             /* Inner loop uses 64 flops */
 439         }
 440
 441         /* End of innermost loop */
 442
 443         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 444                                               f+i_coord_offset,fshift+i_shift_offset);
 445
 446         ggid                        = gid[iidx];
 447         /* Update potential energies */
 448         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 449         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 450
 451         /* Increment number of inner iterations */
 452         inneriter                  += j_index_end - j_index_start;
 453
 454         /* Outer loop uses 9 flops */
 455     }
 456
 457     /* Increment number of outer iterations */
 458     outeriter        += nri;
 459
 460     /* Update outer/inner flops */
 461
 462     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*64);
 463 }
 464 /*
 465  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJEwSh_GeomP1P1_F_avx_128_fma_single
 466  * Electrostatics interaction: Ewald
 467  * VdW interaction:            LJEwald
 468  * Geometry:                   Particle-Particle
 469  * Calculate force/pot:        Force
 470  */
 471 void
 472 nb_kernel_ElecEwSh_VdwLJEwSh_GeomP1P1_F_avx_128_fma_single
 473                     (t_nblist                    * gmx_restrict       nlist,
 474                      rvec                        * gmx_restrict          xx,
 475                      rvec                        * gmx_restrict          ff,
 476                      t_forcerec                  * gmx_restrict          fr,
 477                      t_mdatoms                   * gmx_restrict     mdatoms,
 478                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 479                      t_nrnb                      * gmx_restrict        nrnb)
 480 {
 481     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 482      * just 0 for non-waters.
 483      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 484      * jnr indices corresponding to data put in the four positions in the SIMD register.
 485      */
 486     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 487     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 488     int              jnrA,jnrB,jnrC,jnrD;
 489     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 490     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 491     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 492     real             rcutoff_scalar;
 493     real             *shiftvec,*fshift,*x,*f;
 494     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 495     real             scratch[4*DIM];
 496     __m128           fscal,rcutoff,rcutoff2,jidxall;
 497     int              vdwioffset0;
 498     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 499     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 500     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 501     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 502     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 503     real             *charge;
 504     int              nvdwtype;
 505     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 506     int              *vdwtype;
 507     real             *vdwparam;
 508     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 509     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 510     __m128           c6grid_00;
 511     real             *vdwgridparam;
 512     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 513     __m128           one_half = _mm_set1_ps(0.5);
 514     __m128           minus_one = _mm_set1_ps(-1.0);
 515     __m128i          ewitab;
 516     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 517     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 518     real             *ewtab;
 519     __m128           dummy_mask,cutoff_mask;
 520     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 521     __m128           one     = _mm_set1_ps(1.0);
 522     __m128           two     = _mm_set1_ps(2.0);
 523     x                = xx[0];
 524     f                = ff[0];
 525
 526     nri              = nlist->nri;
 527     iinr             = nlist->iinr;
 528     jindex           = nlist->jindex;
 529     jjnr             = nlist->jjnr;
 530     shiftidx         = nlist->shift;
 531     gid              = nlist->gid;
 532     shiftvec         = fr->shift_vec[0];
 533     fshift           = fr->fshift[0];
 534     facel            = _mm_set1_ps(fr->epsfac);
 535     charge           = mdatoms->chargeA;
 536     nvdwtype         = fr->ntype;
 537     vdwparam         = fr->nbfp;
 538     vdwtype          = mdatoms->typeA;
 539     vdwgridparam     = fr->ljpme_c6grid;
 540     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 541     ewclj            = _mm_set1_ps(fr->ewaldcoeff_lj);
 542     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 543
 544     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 545     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 546     beta2            = _mm_mul_ps(beta,beta);
 547     beta3            = _mm_mul_ps(beta,beta2);
 548     ewtab            = fr->ic->tabq_coul_F;
 549     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 550     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 551
 552     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 553     rcutoff_scalar   = fr->rcoulomb;
 554     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 555     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 556
 557     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 558     rvdw             = _mm_set1_ps(fr->rvdw);
 559
 560     /* Avoid stupid compiler warnings */
 561     jnrA = jnrB = jnrC = jnrD = 0;
 562     j_coord_offsetA = 0;
 563     j_coord_offsetB = 0;
 564     j_coord_offsetC = 0;
 565     j_coord_offsetD = 0;
 566
 567     outeriter        = 0;
 568     inneriter        = 0;
 569
 570     for(iidx=0;iidx<4*DIM;iidx++)
 571     {
 572         scratch[iidx] = 0.0;
 573     }
 574
 575     /* Start outer loop over neighborlists */
 576     for(iidx=0; iidx<nri; iidx++)
 577     {
 578         /* Load shift vector for this list */
 579         i_shift_offset   = DIM*shiftidx[iidx];
 580
 581         /* Load limits for loop over neighbors */
 582         j_index_start    = jindex[iidx];
 583         j_index_end      = jindex[iidx+1];
 584
 585         /* Get outer coordinate index */
 586         inr              = iinr[iidx];
 587         i_coord_offset   = DIM*inr;
 588
 589         /* Load i particle coords and add shift vector */
 590         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 591
 592         fix0             = _mm_setzero_ps();
 593         fiy0             = _mm_setzero_ps();
 594         fiz0             = _mm_setzero_ps();
 595
 596         /* Load parameters for i particles */
 597         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 598         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 599
 600         /* Start inner kernel loop */
 601         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 602         {
 603
 604             /* Get j neighbor index, and coordinate index */
 605             jnrA             = jjnr[jidx];
 606             jnrB             = jjnr[jidx+1];
 607             jnrC             = jjnr[jidx+2];
 608             jnrD             = jjnr[jidx+3];
 609             j_coord_offsetA  = DIM*jnrA;
 610             j_coord_offsetB  = DIM*jnrB;
 611             j_coord_offsetC  = DIM*jnrC;
 612             j_coord_offsetD  = DIM*jnrD;
 613
 614             /* load j atom coordinates */
 615             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 616                                               x+j_coord_offsetC,x+j_coord_offsetD,
 617                                               &jx0,&jy0,&jz0);
 618
 619             /* Calculate displacement vector */
 620             dx00             = _mm_sub_ps(ix0,jx0);
 621             dy00             = _mm_sub_ps(iy0,jy0);
 622             dz00             = _mm_sub_ps(iz0,jz0);
 623
 624             /* Calculate squared distance and things based on it */
 625             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 626
 627             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 628
 629             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 630
 631             /* Load parameters for j particles */
 632             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 633                                                               charge+jnrC+0,charge+jnrD+0);
 634             vdwjidx0A        = 2*vdwtype[jnrA+0];
 635             vdwjidx0B        = 2*vdwtype[jnrB+0];
 636             vdwjidx0C        = 2*vdwtype[jnrC+0];
 637             vdwjidx0D        = 2*vdwtype[jnrD+0];
 638
 639             /**************************
 640              * CALCULATE INTERACTIONS *
 641              **************************/
 642
 643             if (gmx_mm_any_lt(rsq00,rcutoff2))
 644             {
 645
 646             r00              = _mm_mul_ps(rsq00,rinv00);
 647
 648             /* Compute parameters for interactions between i and j atoms */
 649             qq00             = _mm_mul_ps(iq0,jq0);
 650             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 651                                          vdwparam+vdwioffset0+vdwjidx0B,
 652                                          vdwparam+vdwioffset0+vdwjidx0C,
 653                                          vdwparam+vdwioffset0+vdwjidx0D,
 654                                          &c6_00,&c12_00);
 655
 656             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 657                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 658                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 659                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 660
 661             /* EWALD ELECTROSTATICS */
 662
 663             /* Analytical PME correction */
 664             zeta2            = _mm_mul_ps(beta2,rsq00);
 665             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 666             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 667             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 668             felec            = _mm_mul_ps(qq00,felec);
 669
 670             /* Analytical LJ-PME */
 671             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 672             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 673             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 674             exponent         = gmx_simd_exp_r(ewcljrsq);
 675             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 676             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 677             /* f6A = 6 * C6grid * (1 - poly) */
 678             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
 679             /* f6B = C6grid * exponent * beta^6 */
 680             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
 681             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 682             fvdw              = _mm_mul_ps(_mm_macc_ps(_mm_msub_ps(c12_00,rinvsix,_mm_sub_ps(c6_00,f6A)),rinvsix,f6B),rinvsq00);
 683
 684             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 685
 686             fscal            = _mm_add_ps(felec,fvdw);
 687
 688             fscal            = _mm_and_ps(fscal,cutoff_mask);
 689
 690              /* Update vectorial force */
 691             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 692             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 693             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 694
 695             fjptrA             = f+j_coord_offsetA;
 696             fjptrB             = f+j_coord_offsetB;
 697             fjptrC             = f+j_coord_offsetC;
 698             fjptrD             = f+j_coord_offsetD;
 699             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 700                                                    _mm_mul_ps(dx00,fscal),
 701                                                    _mm_mul_ps(dy00,fscal),
 702                                                    _mm_mul_ps(dz00,fscal));
 703
 704             }
 705
 706             /* Inner loop uses 52 flops */
 707         }
 708
 709         if(jidx<j_index_end)
 710         {
 711
 712             /* Get j neighbor index, and coordinate index */
 713             jnrlistA         = jjnr[jidx];
 714             jnrlistB         = jjnr[jidx+1];
 715             jnrlistC         = jjnr[jidx+2];
 716             jnrlistD         = jjnr[jidx+3];
 717             /* Sign of each element will be negative for non-real atoms.
 718              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 719              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 720              */
 721             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 722             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 723             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 724             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 725             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 726             j_coord_offsetA  = DIM*jnrA;
 727             j_coord_offsetB  = DIM*jnrB;
 728             j_coord_offsetC  = DIM*jnrC;
 729             j_coord_offsetD  = DIM*jnrD;
 730
 731             /* load j atom coordinates */
 732             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 733                                               x+j_coord_offsetC,x+j_coord_offsetD,
 734                                               &jx0,&jy0,&jz0);
 735
 736             /* Calculate displacement vector */
 737             dx00             = _mm_sub_ps(ix0,jx0);
 738             dy00             = _mm_sub_ps(iy0,jy0);
 739             dz00             = _mm_sub_ps(iz0,jz0);
 740
 741             /* Calculate squared distance and things based on it */
 742             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 743
 744             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 745
 746             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 747
 748             /* Load parameters for j particles */
 749             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 750                                                               charge+jnrC+0,charge+jnrD+0);
 751             vdwjidx0A        = 2*vdwtype[jnrA+0];
 752             vdwjidx0B        = 2*vdwtype[jnrB+0];
 753             vdwjidx0C        = 2*vdwtype[jnrC+0];
 754             vdwjidx0D        = 2*vdwtype[jnrD+0];
 755
 756             /**************************
 757              * CALCULATE INTERACTIONS *
 758              **************************/
 759
 760             if (gmx_mm_any_lt(rsq00,rcutoff2))
 761             {
 762
 763             r00              = _mm_mul_ps(rsq00,rinv00);
 764             r00              = _mm_andnot_ps(dummy_mask,r00);
 765
 766             /* Compute parameters for interactions between i and j atoms */
 767             qq00             = _mm_mul_ps(iq0,jq0);
 768             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 769                                          vdwparam+vdwioffset0+vdwjidx0B,
 770                                          vdwparam+vdwioffset0+vdwjidx0C,
 771                                          vdwparam+vdwioffset0+vdwjidx0D,
 772                                          &c6_00,&c12_00);
 773
 774             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 775                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 776                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 777                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 778
 779             /* EWALD ELECTROSTATICS */
 780
 781             /* Analytical PME correction */
 782             zeta2            = _mm_mul_ps(beta2,rsq00);
 783             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 784             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 785             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 786             felec            = _mm_mul_ps(qq00,felec);
 787
 788             /* Analytical LJ-PME */
 789             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 790             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 791             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 792             exponent         = gmx_simd_exp_r(ewcljrsq);
 793             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 794             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 795             /* f6A = 6 * C6grid * (1 - poly) */
 796             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
 797             /* f6B = C6grid * exponent * beta^6 */
 798             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
 799             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 800             fvdw              = _mm_mul_ps(_mm_macc_ps(_mm_msub_ps(c12_00,rinvsix,_mm_sub_ps(c6_00,f6A)),rinvsix,f6B),rinvsq00);
 801
 802             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 803
 804             fscal            = _mm_add_ps(felec,fvdw);
 805
 806             fscal            = _mm_and_ps(fscal,cutoff_mask);
 807
 808             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 809
 810              /* Update vectorial force */
 811             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 812             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 813             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 814
 815             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 816             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 817             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 818             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 819             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 820                                                    _mm_mul_ps(dx00,fscal),
 821                                                    _mm_mul_ps(dy00,fscal),
 822                                                    _mm_mul_ps(dz00,fscal));
 823
 824             }
 825
 826             /* Inner loop uses 53 flops */
 827         }
 828
 829         /* End of innermost loop */
 830
 831         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 832                                               f+i_coord_offset,fshift+i_shift_offset);
 833
 834         /* Increment number of inner iterations */
 835         inneriter                  += j_index_end - j_index_start;
 836
 837         /* Outer loop uses 7 flops */
 838     }
 839
 840     /* Increment number of outer iterations */
 841     outeriter        += nri;
 842
 843     /* Update outer/inner flops */
 844
 845     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*53);
 846 }