src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_avx_128_fma_double.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, 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_double kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_128_fma_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_VF_avx_128_fma_double
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            CubicSplineTable
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_VF_avx_128_fma_double
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB;
  74     int              j_coord_offsetA,j_coord_offsetB;
  75     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  76     real             rcutoff_scalar;
  77     real             *shiftvec,*fshift,*x,*f;
  78     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  79     int              vdwioffset0;
  80     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  81     int              vdwjidx0A,vdwjidx0B;
  82     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  83     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  84     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  85     real             *charge;
  86     int              nvdwtype;
  87     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  88     int              *vdwtype;
  89     real             *vdwparam;
  90     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
  91     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
  92     __m128i          vfitab;
  93     __m128i          ifour       = _mm_set1_epi32(4);
  94     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
  95     real             *vftab;
  96     __m128d          dummy_mask,cutoff_mask;
  97     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
  98     __m128d          one     = _mm_set1_pd(1.0);
  99     __m128d          two     = _mm_set1_pd(2.0);
 100     x                = xx[0];
 101     f                = ff[0];
 102
 103     nri              = nlist->nri;
 104     iinr             = nlist->iinr;
 105     jindex           = nlist->jindex;
 106     jjnr             = nlist->jjnr;
 107     shiftidx         = nlist->shift;
 108     gid              = nlist->gid;
 109     shiftvec         = fr->shift_vec[0];
 110     fshift           = fr->fshift[0];
 111     facel            = _mm_set1_pd(fr->ic->epsfac);
 112     charge           = mdatoms->chargeA;
 113     krf              = _mm_set1_pd(fr->ic->k_rf);
 114     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 115     crf              = _mm_set1_pd(fr->ic->c_rf);
 116     nvdwtype         = fr->ntype;
 117     vdwparam         = fr->nbfp;
 118     vdwtype          = mdatoms->typeA;
 119
 120     vftab            = kernel_data->table_vdw->data;
 121     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
 122
 123     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 124     rcutoff_scalar   = fr->ic->rcoulomb;
 125     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 126     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 127
 128     /* Avoid stupid compiler warnings */
 129     jnrA = jnrB = 0;
 130     j_coord_offsetA = 0;
 131     j_coord_offsetB = 0;
 132
 133     outeriter        = 0;
 134     inneriter        = 0;
 135
 136     /* Start outer loop over neighborlists */
 137     for(iidx=0; iidx<nri; iidx++)
 138     {
 139         /* Load shift vector for this list */
 140         i_shift_offset   = DIM*shiftidx[iidx];
 141
 142         /* Load limits for loop over neighbors */
 143         j_index_start    = jindex[iidx];
 144         j_index_end      = jindex[iidx+1];
 145
 146         /* Get outer coordinate index */
 147         inr              = iinr[iidx];
 148         i_coord_offset   = DIM*inr;
 149
 150         /* Load i particle coords and add shift vector */
 151         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 152
 153         fix0             = _mm_setzero_pd();
 154         fiy0             = _mm_setzero_pd();
 155         fiz0             = _mm_setzero_pd();
 156
 157         /* Load parameters for i particles */
 158         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 159         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 160
 161         /* Reset potential sums */
 162         velecsum         = _mm_setzero_pd();
 163         vvdwsum          = _mm_setzero_pd();
 164
 165         /* Start inner kernel loop */
 166         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 167         {
 168
 169             /* Get j neighbor index, and coordinate index */
 170             jnrA             = jjnr[jidx];
 171             jnrB             = jjnr[jidx+1];
 172             j_coord_offsetA  = DIM*jnrA;
 173             j_coord_offsetB  = DIM*jnrB;
 174
 175             /* load j atom coordinates */
 176             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 177                                               &jx0,&jy0,&jz0);
 178
 179             /* Calculate displacement vector */
 180             dx00             = _mm_sub_pd(ix0,jx0);
 181             dy00             = _mm_sub_pd(iy0,jy0);
 182             dz00             = _mm_sub_pd(iz0,jz0);
 183
 184             /* Calculate squared distance and things based on it */
 185             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 186
 187             rinv00           = avx128fma_invsqrt_d(rsq00);
 188
 189             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 190
 191             /* Load parameters for j particles */
 192             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 193             vdwjidx0A        = 2*vdwtype[jnrA+0];
 194             vdwjidx0B        = 2*vdwtype[jnrB+0];
 195
 196             /**************************
 197              * CALCULATE INTERACTIONS *
 198              **************************/
 199
 200             if (gmx_mm_any_lt(rsq00,rcutoff2))
 201             {
 202
 203             r00              = _mm_mul_pd(rsq00,rinv00);
 204
 205             /* Compute parameters for interactions between i and j atoms */
 206             qq00             = _mm_mul_pd(iq0,jq0);
 207             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 208                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 209
 210             /* Calculate table index by multiplying r with table scale and truncate to integer */
 211             rt               = _mm_mul_pd(r00,vftabscale);
 212             vfitab           = _mm_cvttpd_epi32(rt);
 213 #ifdef __XOP__
 214             vfeps            = _mm_frcz_pd(rt);
 215 #else
 216             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 217 #endif
 218             twovfeps         = _mm_add_pd(vfeps,vfeps);
 219             vfitab           = _mm_slli_epi32(vfitab,3);
 220
 221             /* REACTION-FIELD ELECTROSTATICS */
 222             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
 223             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 224
 225             /* CUBIC SPLINE TABLE DISPERSION */
 226             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 227             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 228             GMX_MM_TRANSPOSE2_PD(Y,F);
 229             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 230             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
 231             GMX_MM_TRANSPOSE2_PD(G,H);
 232             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 233             VV               = _mm_macc_pd(vfeps,Fp,Y);
 234             vvdw6            = _mm_mul_pd(c6_00,VV);
 235             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 236             fvdw6            = _mm_mul_pd(c6_00,FF);
 237
 238             /* CUBIC SPLINE TABLE REPULSION */
 239             vfitab           = _mm_add_epi32(vfitab,ifour);
 240             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 241             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 242             GMX_MM_TRANSPOSE2_PD(Y,F);
 243             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 244             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
 245             GMX_MM_TRANSPOSE2_PD(G,H);
 246             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 247             VV               = _mm_macc_pd(vfeps,Fp,Y);
 248             vvdw12           = _mm_mul_pd(c12_00,VV);
 249             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 250             fvdw12           = _mm_mul_pd(c12_00,FF);
 251             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 252             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 253
 254             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 255
 256             /* Update potential sum for this i atom from the interaction with this j atom. */
 257             velec            = _mm_and_pd(velec,cutoff_mask);
 258             velecsum         = _mm_add_pd(velecsum,velec);
 259             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 260             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 261
 262             fscal            = _mm_add_pd(felec,fvdw);
 263
 264             fscal            = _mm_and_pd(fscal,cutoff_mask);
 265
 266             /* Update vectorial force */
 267             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 268             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 269             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 270
 271             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 272                                                    _mm_mul_pd(dx00,fscal),
 273                                                    _mm_mul_pd(dy00,fscal),
 274                                                    _mm_mul_pd(dz00,fscal));
 275
 276             }
 277
 278             /* Inner loop uses 75 flops */
 279         }
 280
 281         if(jidx<j_index_end)
 282         {
 283
 284             jnrA             = jjnr[jidx];
 285             j_coord_offsetA  = DIM*jnrA;
 286
 287             /* load j atom coordinates */
 288             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 289                                               &jx0,&jy0,&jz0);
 290
 291             /* Calculate displacement vector */
 292             dx00             = _mm_sub_pd(ix0,jx0);
 293             dy00             = _mm_sub_pd(iy0,jy0);
 294             dz00             = _mm_sub_pd(iz0,jz0);
 295
 296             /* Calculate squared distance and things based on it */
 297             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 298
 299             rinv00           = avx128fma_invsqrt_d(rsq00);
 300
 301             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 302
 303             /* Load parameters for j particles */
 304             jq0              = _mm_load_sd(charge+jnrA+0);
 305             vdwjidx0A        = 2*vdwtype[jnrA+0];
 306
 307             /**************************
 308              * CALCULATE INTERACTIONS *
 309              **************************/
 310
 311             if (gmx_mm_any_lt(rsq00,rcutoff2))
 312             {
 313
 314             r00              = _mm_mul_pd(rsq00,rinv00);
 315
 316             /* Compute parameters for interactions between i and j atoms */
 317             qq00             = _mm_mul_pd(iq0,jq0);
 318             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 319
 320             /* Calculate table index by multiplying r with table scale and truncate to integer */
 321             rt               = _mm_mul_pd(r00,vftabscale);
 322             vfitab           = _mm_cvttpd_epi32(rt);
 323 #ifdef __XOP__
 324             vfeps            = _mm_frcz_pd(rt);
 325 #else
 326             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 327 #endif
 328             twovfeps         = _mm_add_pd(vfeps,vfeps);
 329             vfitab           = _mm_slli_epi32(vfitab,3);
 330
 331             /* REACTION-FIELD ELECTROSTATICS */
 332             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
 333             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 334
 335             /* CUBIC SPLINE TABLE DISPERSION */
 336             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 337             F                = _mm_setzero_pd();
 338             GMX_MM_TRANSPOSE2_PD(Y,F);
 339             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 340             H                = _mm_setzero_pd();
 341             GMX_MM_TRANSPOSE2_PD(G,H);
 342             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 343             VV               = _mm_macc_pd(vfeps,Fp,Y);
 344             vvdw6            = _mm_mul_pd(c6_00,VV);
 345             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 346             fvdw6            = _mm_mul_pd(c6_00,FF);
 347
 348             /* CUBIC SPLINE TABLE REPULSION */
 349             vfitab           = _mm_add_epi32(vfitab,ifour);
 350             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 351             F                = _mm_setzero_pd();
 352             GMX_MM_TRANSPOSE2_PD(Y,F);
 353             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 354             H                = _mm_setzero_pd();
 355             GMX_MM_TRANSPOSE2_PD(G,H);
 356             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 357             VV               = _mm_macc_pd(vfeps,Fp,Y);
 358             vvdw12           = _mm_mul_pd(c12_00,VV);
 359             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 360             fvdw12           = _mm_mul_pd(c12_00,FF);
 361             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 362             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 363
 364             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 365
 366             /* Update potential sum for this i atom from the interaction with this j atom. */
 367             velec            = _mm_and_pd(velec,cutoff_mask);
 368             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 369             velecsum         = _mm_add_pd(velecsum,velec);
 370             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 371             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 372             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 373
 374             fscal            = _mm_add_pd(felec,fvdw);
 375
 376             fscal            = _mm_and_pd(fscal,cutoff_mask);
 377
 378             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 379
 380             /* Update vectorial force */
 381             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 382             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 383             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 384
 385             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 386                                                    _mm_mul_pd(dx00,fscal),
 387                                                    _mm_mul_pd(dy00,fscal),
 388                                                    _mm_mul_pd(dz00,fscal));
 389
 390             }
 391
 392             /* Inner loop uses 75 flops */
 393         }
 394
 395         /* End of innermost loop */
 396
 397         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 398                                               f+i_coord_offset,fshift+i_shift_offset);
 399
 400         ggid                        = gid[iidx];
 401         /* Update potential energies */
 402         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 403         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 404
 405         /* Increment number of inner iterations */
 406         inneriter                  += j_index_end - j_index_start;
 407
 408         /* Outer loop uses 9 flops */
 409     }
 410
 411     /* Increment number of outer iterations */
 412     outeriter        += nri;
 413
 414     /* Update outer/inner flops */
 415
 416     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*75);
 417 }
 418 /*
 419  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_F_avx_128_fma_double
 420  * Electrostatics interaction: ReactionField
 421  * VdW interaction:            CubicSplineTable
 422  * Geometry:                   Particle-Particle
 423  * Calculate force/pot:        Force
 424  */
 425 void
 426 nb_kernel_ElecRFCut_VdwCSTab_GeomP1P1_F_avx_128_fma_double
 427                     (t_nblist                    * gmx_restrict       nlist,
 428                      rvec                        * gmx_restrict          xx,
 429                      rvec                        * gmx_restrict          ff,
 430                      struct t_forcerec           * gmx_restrict          fr,
 431                      t_mdatoms                   * gmx_restrict     mdatoms,
 432                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 433                      t_nrnb                      * gmx_restrict        nrnb)
 434 {
 435     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 436      * just 0 for non-waters.
 437      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 438      * jnr indices corresponding to data put in the four positions in the SIMD register.
 439      */
 440     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 441     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 442     int              jnrA,jnrB;
 443     int              j_coord_offsetA,j_coord_offsetB;
 444     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 445     real             rcutoff_scalar;
 446     real             *shiftvec,*fshift,*x,*f;
 447     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 448     int              vdwioffset0;
 449     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 450     int              vdwjidx0A,vdwjidx0B;
 451     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 452     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 453     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 454     real             *charge;
 455     int              nvdwtype;
 456     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 457     int              *vdwtype;
 458     real             *vdwparam;
 459     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 460     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 461     __m128i          vfitab;
 462     __m128i          ifour       = _mm_set1_epi32(4);
 463     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
 464     real             *vftab;
 465     __m128d          dummy_mask,cutoff_mask;
 466     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 467     __m128d          one     = _mm_set1_pd(1.0);
 468     __m128d          two     = _mm_set1_pd(2.0);
 469     x                = xx[0];
 470     f                = ff[0];
 471
 472     nri              = nlist->nri;
 473     iinr             = nlist->iinr;
 474     jindex           = nlist->jindex;
 475     jjnr             = nlist->jjnr;
 476     shiftidx         = nlist->shift;
 477     gid              = nlist->gid;
 478     shiftvec         = fr->shift_vec[0];
 479     fshift           = fr->fshift[0];
 480     facel            = _mm_set1_pd(fr->ic->epsfac);
 481     charge           = mdatoms->chargeA;
 482     krf              = _mm_set1_pd(fr->ic->k_rf);
 483     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 484     crf              = _mm_set1_pd(fr->ic->c_rf);
 485     nvdwtype         = fr->ntype;
 486     vdwparam         = fr->nbfp;
 487     vdwtype          = mdatoms->typeA;
 488
 489     vftab            = kernel_data->table_vdw->data;
 490     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
 491
 492     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 493     rcutoff_scalar   = fr->ic->rcoulomb;
 494     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 495     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 496
 497     /* Avoid stupid compiler warnings */
 498     jnrA = jnrB = 0;
 499     j_coord_offsetA = 0;
 500     j_coord_offsetB = 0;
 501
 502     outeriter        = 0;
 503     inneriter        = 0;
 504
 505     /* Start outer loop over neighborlists */
 506     for(iidx=0; iidx<nri; iidx++)
 507     {
 508         /* Load shift vector for this list */
 509         i_shift_offset   = DIM*shiftidx[iidx];
 510
 511         /* Load limits for loop over neighbors */
 512         j_index_start    = jindex[iidx];
 513         j_index_end      = jindex[iidx+1];
 514
 515         /* Get outer coordinate index */
 516         inr              = iinr[iidx];
 517         i_coord_offset   = DIM*inr;
 518
 519         /* Load i particle coords and add shift vector */
 520         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 521
 522         fix0             = _mm_setzero_pd();
 523         fiy0             = _mm_setzero_pd();
 524         fiz0             = _mm_setzero_pd();
 525
 526         /* Load parameters for i particles */
 527         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 528         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 529
 530         /* Start inner kernel loop */
 531         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 532         {
 533
 534             /* Get j neighbor index, and coordinate index */
 535             jnrA             = jjnr[jidx];
 536             jnrB             = jjnr[jidx+1];
 537             j_coord_offsetA  = DIM*jnrA;
 538             j_coord_offsetB  = DIM*jnrB;
 539
 540             /* load j atom coordinates */
 541             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 542                                               &jx0,&jy0,&jz0);
 543
 544             /* Calculate displacement vector */
 545             dx00             = _mm_sub_pd(ix0,jx0);
 546             dy00             = _mm_sub_pd(iy0,jy0);
 547             dz00             = _mm_sub_pd(iz0,jz0);
 548
 549             /* Calculate squared distance and things based on it */
 550             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 551
 552             rinv00           = avx128fma_invsqrt_d(rsq00);
 553
 554             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 555
 556             /* Load parameters for j particles */
 557             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 558             vdwjidx0A        = 2*vdwtype[jnrA+0];
 559             vdwjidx0B        = 2*vdwtype[jnrB+0];
 560
 561             /**************************
 562              * CALCULATE INTERACTIONS *
 563              **************************/
 564
 565             if (gmx_mm_any_lt(rsq00,rcutoff2))
 566             {
 567
 568             r00              = _mm_mul_pd(rsq00,rinv00);
 569
 570             /* Compute parameters for interactions between i and j atoms */
 571             qq00             = _mm_mul_pd(iq0,jq0);
 572             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 573                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 574
 575             /* Calculate table index by multiplying r with table scale and truncate to integer */
 576             rt               = _mm_mul_pd(r00,vftabscale);
 577             vfitab           = _mm_cvttpd_epi32(rt);
 578 #ifdef __XOP__
 579             vfeps            = _mm_frcz_pd(rt);
 580 #else
 581             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 582 #endif
 583             twovfeps         = _mm_add_pd(vfeps,vfeps);
 584             vfitab           = _mm_slli_epi32(vfitab,3);
 585
 586             /* REACTION-FIELD ELECTROSTATICS */
 587             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 588
 589             /* CUBIC SPLINE TABLE DISPERSION */
 590             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 591             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 592             GMX_MM_TRANSPOSE2_PD(Y,F);
 593             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 594             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
 595             GMX_MM_TRANSPOSE2_PD(G,H);
 596             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 597             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 598             fvdw6            = _mm_mul_pd(c6_00,FF);
 599
 600             /* CUBIC SPLINE TABLE REPULSION */
 601             vfitab           = _mm_add_epi32(vfitab,ifour);
 602             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 603             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 604             GMX_MM_TRANSPOSE2_PD(Y,F);
 605             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 606             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
 607             GMX_MM_TRANSPOSE2_PD(G,H);
 608             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 609             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 610             fvdw12           = _mm_mul_pd(c12_00,FF);
 611             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 612
 613             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 614
 615             fscal            = _mm_add_pd(felec,fvdw);
 616
 617             fscal            = _mm_and_pd(fscal,cutoff_mask);
 618
 619             /* Update vectorial force */
 620             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 621             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 622             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 623
 624             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 625                                                    _mm_mul_pd(dx00,fscal),
 626                                                    _mm_mul_pd(dy00,fscal),
 627                                                    _mm_mul_pd(dz00,fscal));
 628
 629             }
 630
 631             /* Inner loop uses 60 flops */
 632         }
 633
 634         if(jidx<j_index_end)
 635         {
 636
 637             jnrA             = jjnr[jidx];
 638             j_coord_offsetA  = DIM*jnrA;
 639
 640             /* load j atom coordinates */
 641             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 642                                               &jx0,&jy0,&jz0);
 643
 644             /* Calculate displacement vector */
 645             dx00             = _mm_sub_pd(ix0,jx0);
 646             dy00             = _mm_sub_pd(iy0,jy0);
 647             dz00             = _mm_sub_pd(iz0,jz0);
 648
 649             /* Calculate squared distance and things based on it */
 650             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 651
 652             rinv00           = avx128fma_invsqrt_d(rsq00);
 653
 654             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 655
 656             /* Load parameters for j particles */
 657             jq0              = _mm_load_sd(charge+jnrA+0);
 658             vdwjidx0A        = 2*vdwtype[jnrA+0];
 659
 660             /**************************
 661              * CALCULATE INTERACTIONS *
 662              **************************/
 663
 664             if (gmx_mm_any_lt(rsq00,rcutoff2))
 665             {
 666
 667             r00              = _mm_mul_pd(rsq00,rinv00);
 668
 669             /* Compute parameters for interactions between i and j atoms */
 670             qq00             = _mm_mul_pd(iq0,jq0);
 671             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 672
 673             /* Calculate table index by multiplying r with table scale and truncate to integer */
 674             rt               = _mm_mul_pd(r00,vftabscale);
 675             vfitab           = _mm_cvttpd_epi32(rt);
 676 #ifdef __XOP__
 677             vfeps            = _mm_frcz_pd(rt);
 678 #else
 679             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 680 #endif
 681             twovfeps         = _mm_add_pd(vfeps,vfeps);
 682             vfitab           = _mm_slli_epi32(vfitab,3);
 683
 684             /* REACTION-FIELD ELECTROSTATICS */
 685             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 686
 687             /* CUBIC SPLINE TABLE DISPERSION */
 688             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 689             F                = _mm_setzero_pd();
 690             GMX_MM_TRANSPOSE2_PD(Y,F);
 691             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 692             H                = _mm_setzero_pd();
 693             GMX_MM_TRANSPOSE2_PD(G,H);
 694             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 695             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 696             fvdw6            = _mm_mul_pd(c6_00,FF);
 697
 698             /* CUBIC SPLINE TABLE REPULSION */
 699             vfitab           = _mm_add_epi32(vfitab,ifour);
 700             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 701             F                = _mm_setzero_pd();
 702             GMX_MM_TRANSPOSE2_PD(Y,F);
 703             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 704             H                = _mm_setzero_pd();
 705             GMX_MM_TRANSPOSE2_PD(G,H);
 706             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 707             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 708             fvdw12           = _mm_mul_pd(c12_00,FF);
 709             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 710
 711             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 712
 713             fscal            = _mm_add_pd(felec,fvdw);
 714
 715             fscal            = _mm_and_pd(fscal,cutoff_mask);
 716
 717             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 718
 719             /* Update vectorial force */
 720             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 721             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 722             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 723
 724             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 725                                                    _mm_mul_pd(dx00,fscal),
 726                                                    _mm_mul_pd(dy00,fscal),
 727                                                    _mm_mul_pd(dz00,fscal));
 728
 729             }
 730
 731             /* Inner loop uses 60 flops */
 732         }
 733
 734         /* End of innermost loop */
 735
 736         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 737                                               f+i_coord_offset,fshift+i_shift_offset);
 738
 739         /* Increment number of inner iterations */
 740         inneriter                  += j_index_end - j_index_start;
 741
 742         /* Outer loop uses 7 flops */
 743     }
 744
 745     /* Increment number of outer iterations */
 746     outeriter        += nri;
 747
 748     /* Update outer/inner flops */
 749
 750     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*60);
 751 }