src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecEwSh_VdwLJEwSh_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_ElecEwSh_VdwLJEwSh_GeomP1P1_VF_avx_128_fma_double
  51  * Electrostatics interaction: Ewald
  52  * VdW interaction:            LJEwald
  53  * Geometry:                   Particle-Particle
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecEwSh_VdwLJEwSh_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     __m128d           c6grid_00;
  93     real             *vdwgridparam;
  94     __m128d           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
  95     __m128d           one_half  = _mm_set1_pd(0.5);
  96     __m128d           minus_one = _mm_set1_pd(-1.0);
  97     __m128i          ewitab;
  98     __m128d          ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  99     real             *ewtab;
 100     __m128d          dummy_mask,cutoff_mask;
 101     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 102     __m128d          one     = _mm_set1_pd(1.0);
 103     __m128d          two     = _mm_set1_pd(2.0);
 104     x                = xx[0];
 105     f                = ff[0];
 106
 107     nri              = nlist->nri;
 108     iinr             = nlist->iinr;
 109     jindex           = nlist->jindex;
 110     jjnr             = nlist->jjnr;
 111     shiftidx         = nlist->shift;
 112     gid              = nlist->gid;
 113     shiftvec         = fr->shift_vec[0];
 114     fshift           = fr->fshift[0];
 115     facel            = _mm_set1_pd(fr->ic->epsfac);
 116     charge           = mdatoms->chargeA;
 117     nvdwtype         = fr->ntype;
 118     vdwparam         = fr->nbfp;
 119     vdwtype          = mdatoms->typeA;
 120     vdwgridparam     = fr->ljpme_c6grid;
 121     sh_lj_ewald      = _mm_set1_pd(fr->ic->sh_lj_ewald);
 122     ewclj            = _mm_set1_pd(fr->ic->ewaldcoeff_lj);
 123     ewclj2           = _mm_mul_pd(minus_one,_mm_mul_pd(ewclj,ewclj));
 124
 125     sh_ewald         = _mm_set1_pd(fr->ic->sh_ewald);
 126     ewtab            = fr->ic->tabq_coul_FDV0;
 127     ewtabscale       = _mm_set1_pd(fr->ic->tabq_scale);
 128     ewtabhalfspace   = _mm_set1_pd(0.5/fr->ic->tabq_scale);
 129
 130     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 131     rcutoff_scalar   = fr->ic->rcoulomb;
 132     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 133     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 134
 135     sh_vdw_invrcut6  = _mm_set1_pd(fr->ic->sh_invrc6);
 136     rvdw             = _mm_set1_pd(fr->ic->rvdw);
 137
 138     /* Avoid stupid compiler warnings */
 139     jnrA = jnrB = 0;
 140     j_coord_offsetA = 0;
 141     j_coord_offsetB = 0;
 142
 143     outeriter        = 0;
 144     inneriter        = 0;
 145
 146     /* Start outer loop over neighborlists */
 147     for(iidx=0; iidx<nri; iidx++)
 148     {
 149         /* Load shift vector for this list */
 150         i_shift_offset   = DIM*shiftidx[iidx];
 151
 152         /* Load limits for loop over neighbors */
 153         j_index_start    = jindex[iidx];
 154         j_index_end      = jindex[iidx+1];
 155
 156         /* Get outer coordinate index */
 157         inr              = iinr[iidx];
 158         i_coord_offset   = DIM*inr;
 159
 160         /* Load i particle coords and add shift vector */
 161         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 162
 163         fix0             = _mm_setzero_pd();
 164         fiy0             = _mm_setzero_pd();
 165         fiz0             = _mm_setzero_pd();
 166
 167         /* Load parameters for i particles */
 168         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 169         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 170
 171         /* Reset potential sums */
 172         velecsum         = _mm_setzero_pd();
 173         vvdwsum          = _mm_setzero_pd();
 174
 175         /* Start inner kernel loop */
 176         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 177         {
 178
 179             /* Get j neighbor index, and coordinate index */
 180             jnrA             = jjnr[jidx];
 181             jnrB             = jjnr[jidx+1];
 182             j_coord_offsetA  = DIM*jnrA;
 183             j_coord_offsetB  = DIM*jnrB;
 184
 185             /* load j atom coordinates */
 186             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 187                                               &jx0,&jy0,&jz0);
 188
 189             /* Calculate displacement vector */
 190             dx00             = _mm_sub_pd(ix0,jx0);
 191             dy00             = _mm_sub_pd(iy0,jy0);
 192             dz00             = _mm_sub_pd(iz0,jz0);
 193
 194             /* Calculate squared distance and things based on it */
 195             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 196
 197             rinv00           = avx128fma_invsqrt_d(rsq00);
 198
 199             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 200
 201             /* Load parameters for j particles */
 202             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 203             vdwjidx0A        = 2*vdwtype[jnrA+0];
 204             vdwjidx0B        = 2*vdwtype[jnrB+0];
 205
 206             /**************************
 207              * CALCULATE INTERACTIONS *
 208              **************************/
 209
 210             if (gmx_mm_any_lt(rsq00,rcutoff2))
 211             {
 212
 213             r00              = _mm_mul_pd(rsq00,rinv00);
 214
 215             /* Compute parameters for interactions between i and j atoms */
 216             qq00             = _mm_mul_pd(iq0,jq0);
 217             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 218                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 219             c6grid_00       = gmx_mm_load_2real_swizzle_pd(vdwgridparam+vdwioffset0+vdwjidx0A,
 220                                                                vdwgridparam+vdwioffset0+vdwjidx0B);
 221
 222             /* EWALD ELECTROSTATICS */
 223
 224             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 225             ewrt             = _mm_mul_pd(r00,ewtabscale);
 226             ewitab           = _mm_cvttpd_epi32(ewrt);
 227 #ifdef __XOP__
 228             eweps            = _mm_frcz_pd(ewrt);
 229 #else
 230             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 231 #endif
 232             twoeweps         = _mm_add_pd(eweps,eweps);
 233             ewitab           = _mm_slli_epi32(ewitab,2);
 234             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 235             ewtabD           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 236             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 237             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 238             ewtabFn          = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,1) +2);
 239             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 240             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 241             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 242             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_sub_pd(rinv00,sh_ewald),velec));
 243             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 244
 245             /* Analytical LJ-PME */
 246             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 247             ewcljrsq         = _mm_mul_pd(ewclj2,rsq00);
 248             ewclj6           = _mm_mul_pd(ewclj2,_mm_mul_pd(ewclj2,ewclj2));
 249             exponent         = avx128fma_exp_d(ewcljrsq);
 250             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 251             poly             = _mm_mul_pd(exponent,_mm_macc_pd(_mm_mul_pd(ewcljrsq,ewcljrsq),one_half,_mm_sub_pd(one,ewcljrsq)));
 252             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 253             vvdw6            = _mm_mul_pd(_mm_macc_pd(-c6grid_00,_mm_sub_pd(one,poly),c6_00),rinvsix);
 254             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 255             vvdw             = _mm_msub_pd(_mm_nmacc_pd(c12_00,_mm_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 256                                _mm_mul_pd(_mm_sub_pd(vvdw6,_mm_macc_pd(c6grid_00,sh_lj_ewald,_mm_mul_pd(c6_00,sh_vdw_invrcut6))),one_sixth));
 257             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 258             fvdw             = _mm_mul_pd(_mm_add_pd(vvdw12,_mm_msub_pd(_mm_mul_pd(c6grid_00,one_sixth),_mm_mul_pd(exponent,ewclj6),vvdw6)),rinvsq00);
 259
 260             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 261
 262             /* Update potential sum for this i atom from the interaction with this j atom. */
 263             velec            = _mm_and_pd(velec,cutoff_mask);
 264             velecsum         = _mm_add_pd(velecsum,velec);
 265             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 266             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 267
 268             fscal            = _mm_add_pd(felec,fvdw);
 269
 270             fscal            = _mm_and_pd(fscal,cutoff_mask);
 271
 272             /* Update vectorial force */
 273             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 274             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 275             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 276
 277             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 278                                                    _mm_mul_pd(dx00,fscal),
 279                                                    _mm_mul_pd(dy00,fscal),
 280                                                    _mm_mul_pd(dz00,fscal));
 281
 282             }
 283
 284             /* Inner loop uses 78 flops */
 285         }
 286
 287         if(jidx<j_index_end)
 288         {
 289
 290             jnrA             = jjnr[jidx];
 291             j_coord_offsetA  = DIM*jnrA;
 292
 293             /* load j atom coordinates */
 294             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 295                                               &jx0,&jy0,&jz0);
 296
 297             /* Calculate displacement vector */
 298             dx00             = _mm_sub_pd(ix0,jx0);
 299             dy00             = _mm_sub_pd(iy0,jy0);
 300             dz00             = _mm_sub_pd(iz0,jz0);
 301
 302             /* Calculate squared distance and things based on it */
 303             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 304
 305             rinv00           = avx128fma_invsqrt_d(rsq00);
 306
 307             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 308
 309             /* Load parameters for j particles */
 310             jq0              = _mm_load_sd(charge+jnrA+0);
 311             vdwjidx0A        = 2*vdwtype[jnrA+0];
 312
 313             /**************************
 314              * CALCULATE INTERACTIONS *
 315              **************************/
 316
 317             if (gmx_mm_any_lt(rsq00,rcutoff2))
 318             {
 319
 320             r00              = _mm_mul_pd(rsq00,rinv00);
 321
 322             /* Compute parameters for interactions between i and j atoms */
 323             qq00             = _mm_mul_pd(iq0,jq0);
 324             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 325             c6grid_00       = gmx_mm_load_1real_pd(vdwgridparam+vdwioffset0+vdwjidx0A);
 326
 327             /* EWALD ELECTROSTATICS */
 328
 329             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 330             ewrt             = _mm_mul_pd(r00,ewtabscale);
 331             ewitab           = _mm_cvttpd_epi32(ewrt);
 332 #ifdef __XOP__
 333             eweps            = _mm_frcz_pd(ewrt);
 334 #else
 335             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 336 #endif
 337             twoeweps         = _mm_add_pd(eweps,eweps);
 338             ewitab           = _mm_slli_epi32(ewitab,2);
 339             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 340             ewtabD           = _mm_setzero_pd();
 341             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 342             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 343             ewtabFn          = _mm_setzero_pd();
 344             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 345             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 346             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 347             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_sub_pd(rinv00,sh_ewald),velec));
 348             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 349
 350             /* Analytical LJ-PME */
 351             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 352             ewcljrsq         = _mm_mul_pd(ewclj2,rsq00);
 353             ewclj6           = _mm_mul_pd(ewclj2,_mm_mul_pd(ewclj2,ewclj2));
 354             exponent         = avx128fma_exp_d(ewcljrsq);
 355             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 356             poly             = _mm_mul_pd(exponent,_mm_macc_pd(_mm_mul_pd(ewcljrsq,ewcljrsq),one_half,_mm_sub_pd(one,ewcljrsq)));
 357             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 358             vvdw6            = _mm_mul_pd(_mm_macc_pd(-c6grid_00,_mm_sub_pd(one,poly),c6_00),rinvsix);
 359             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 360             vvdw             = _mm_msub_pd(_mm_nmacc_pd(c12_00,_mm_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 361                                _mm_mul_pd(_mm_sub_pd(vvdw6,_mm_macc_pd(c6grid_00,sh_lj_ewald,_mm_mul_pd(c6_00,sh_vdw_invrcut6))),one_sixth));
 362             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 363             fvdw             = _mm_mul_pd(_mm_add_pd(vvdw12,_mm_msub_pd(_mm_mul_pd(c6grid_00,one_sixth),_mm_mul_pd(exponent,ewclj6),vvdw6)),rinvsq00);
 364
 365             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 366
 367             /* Update potential sum for this i atom from the interaction with this j atom. */
 368             velec            = _mm_and_pd(velec,cutoff_mask);
 369             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 370             velecsum         = _mm_add_pd(velecsum,velec);
 371             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 372             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 373             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 374
 375             fscal            = _mm_add_pd(felec,fvdw);
 376
 377             fscal            = _mm_and_pd(fscal,cutoff_mask);
 378
 379             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 380
 381             /* Update vectorial force */
 382             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 383             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 384             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 385
 386             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 387                                                    _mm_mul_pd(dx00,fscal),
 388                                                    _mm_mul_pd(dy00,fscal),
 389                                                    _mm_mul_pd(dz00,fscal));
 390
 391             }
 392
 393             /* Inner loop uses 78 flops */
 394         }
 395
 396         /* End of innermost loop */
 397
 398         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 399                                               f+i_coord_offset,fshift+i_shift_offset);
 400
 401         ggid                        = gid[iidx];
 402         /* Update potential energies */
 403         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 404         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 405
 406         /* Increment number of inner iterations */
 407         inneriter                  += j_index_end - j_index_start;
 408
 409         /* Outer loop uses 9 flops */
 410     }
 411
 412     /* Increment number of outer iterations */
 413     outeriter        += nri;
 414
 415     /* Update outer/inner flops */
 416
 417     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*78);
 418 }
 419 /*
 420  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJEwSh_GeomP1P1_F_avx_128_fma_double
 421  * Electrostatics interaction: Ewald
 422  * VdW interaction:            LJEwald
 423  * Geometry:                   Particle-Particle
 424  * Calculate force/pot:        Force
 425  */
 426 void
 427 nb_kernel_ElecEwSh_VdwLJEwSh_GeomP1P1_F_avx_128_fma_double
 428                     (t_nblist                    * gmx_restrict       nlist,
 429                      rvec                        * gmx_restrict          xx,
 430                      rvec                        * gmx_restrict          ff,
 431                      struct t_forcerec           * gmx_restrict          fr,
 432                      t_mdatoms                   * gmx_restrict     mdatoms,
 433                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 434                      t_nrnb                      * gmx_restrict        nrnb)
 435 {
 436     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 437      * just 0 for non-waters.
 438      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 439      * jnr indices corresponding to data put in the four positions in the SIMD register.
 440      */
 441     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 442     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 443     int              jnrA,jnrB;
 444     int              j_coord_offsetA,j_coord_offsetB;
 445     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 446     real             rcutoff_scalar;
 447     real             *shiftvec,*fshift,*x,*f;
 448     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 449     int              vdwioffset0;
 450     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 451     int              vdwjidx0A,vdwjidx0B;
 452     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 453     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 454     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 455     real             *charge;
 456     int              nvdwtype;
 457     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 458     int              *vdwtype;
 459     real             *vdwparam;
 460     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 461     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 462     __m128d           c6grid_00;
 463     real             *vdwgridparam;
 464     __m128d           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 465     __m128d           one_half  = _mm_set1_pd(0.5);
 466     __m128d           minus_one = _mm_set1_pd(-1.0);
 467     __m128i          ewitab;
 468     __m128d          ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 469     real             *ewtab;
 470     __m128d          dummy_mask,cutoff_mask;
 471     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 472     __m128d          one     = _mm_set1_pd(1.0);
 473     __m128d          two     = _mm_set1_pd(2.0);
 474     x                = xx[0];
 475     f                = ff[0];
 476
 477     nri              = nlist->nri;
 478     iinr             = nlist->iinr;
 479     jindex           = nlist->jindex;
 480     jjnr             = nlist->jjnr;
 481     shiftidx         = nlist->shift;
 482     gid              = nlist->gid;
 483     shiftvec         = fr->shift_vec[0];
 484     fshift           = fr->fshift[0];
 485     facel            = _mm_set1_pd(fr->ic->epsfac);
 486     charge           = mdatoms->chargeA;
 487     nvdwtype         = fr->ntype;
 488     vdwparam         = fr->nbfp;
 489     vdwtype          = mdatoms->typeA;
 490     vdwgridparam     = fr->ljpme_c6grid;
 491     sh_lj_ewald      = _mm_set1_pd(fr->ic->sh_lj_ewald);
 492     ewclj            = _mm_set1_pd(fr->ic->ewaldcoeff_lj);
 493     ewclj2           = _mm_mul_pd(minus_one,_mm_mul_pd(ewclj,ewclj));
 494
 495     sh_ewald         = _mm_set1_pd(fr->ic->sh_ewald);
 496     ewtab            = fr->ic->tabq_coul_F;
 497     ewtabscale       = _mm_set1_pd(fr->ic->tabq_scale);
 498     ewtabhalfspace   = _mm_set1_pd(0.5/fr->ic->tabq_scale);
 499
 500     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 501     rcutoff_scalar   = fr->ic->rcoulomb;
 502     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 503     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 504
 505     sh_vdw_invrcut6  = _mm_set1_pd(fr->ic->sh_invrc6);
 506     rvdw             = _mm_set1_pd(fr->ic->rvdw);
 507
 508     /* Avoid stupid compiler warnings */
 509     jnrA = jnrB = 0;
 510     j_coord_offsetA = 0;
 511     j_coord_offsetB = 0;
 512
 513     outeriter        = 0;
 514     inneriter        = 0;
 515
 516     /* Start outer loop over neighborlists */
 517     for(iidx=0; iidx<nri; iidx++)
 518     {
 519         /* Load shift vector for this list */
 520         i_shift_offset   = DIM*shiftidx[iidx];
 521
 522         /* Load limits for loop over neighbors */
 523         j_index_start    = jindex[iidx];
 524         j_index_end      = jindex[iidx+1];
 525
 526         /* Get outer coordinate index */
 527         inr              = iinr[iidx];
 528         i_coord_offset   = DIM*inr;
 529
 530         /* Load i particle coords and add shift vector */
 531         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 532
 533         fix0             = _mm_setzero_pd();
 534         fiy0             = _mm_setzero_pd();
 535         fiz0             = _mm_setzero_pd();
 536
 537         /* Load parameters for i particles */
 538         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 539         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 540
 541         /* Start inner kernel loop */
 542         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 543         {
 544
 545             /* Get j neighbor index, and coordinate index */
 546             jnrA             = jjnr[jidx];
 547             jnrB             = jjnr[jidx+1];
 548             j_coord_offsetA  = DIM*jnrA;
 549             j_coord_offsetB  = DIM*jnrB;
 550
 551             /* load j atom coordinates */
 552             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 553                                               &jx0,&jy0,&jz0);
 554
 555             /* Calculate displacement vector */
 556             dx00             = _mm_sub_pd(ix0,jx0);
 557             dy00             = _mm_sub_pd(iy0,jy0);
 558             dz00             = _mm_sub_pd(iz0,jz0);
 559
 560             /* Calculate squared distance and things based on it */
 561             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 562
 563             rinv00           = avx128fma_invsqrt_d(rsq00);
 564
 565             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 566
 567             /* Load parameters for j particles */
 568             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 569             vdwjidx0A        = 2*vdwtype[jnrA+0];
 570             vdwjidx0B        = 2*vdwtype[jnrB+0];
 571
 572             /**************************
 573              * CALCULATE INTERACTIONS *
 574              **************************/
 575
 576             if (gmx_mm_any_lt(rsq00,rcutoff2))
 577             {
 578
 579             r00              = _mm_mul_pd(rsq00,rinv00);
 580
 581             /* Compute parameters for interactions between i and j atoms */
 582             qq00             = _mm_mul_pd(iq0,jq0);
 583             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 584                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 585             c6grid_00       = gmx_mm_load_2real_swizzle_pd(vdwgridparam+vdwioffset0+vdwjidx0A,
 586                                                                vdwgridparam+vdwioffset0+vdwjidx0B);
 587
 588             /* EWALD ELECTROSTATICS */
 589
 590             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 591             ewrt             = _mm_mul_pd(r00,ewtabscale);
 592             ewitab           = _mm_cvttpd_epi32(ewrt);
 593 #ifdef __XOP__
 594             eweps            = _mm_frcz_pd(ewrt);
 595 #else
 596             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 597 #endif
 598             twoeweps         = _mm_add_pd(eweps,eweps);
 599             gmx_mm_load_2pair_swizzle_pd(ewtab+_mm_extract_epi32(ewitab,0),ewtab+_mm_extract_epi32(ewitab,1),
 600                                          &ewtabF,&ewtabFn);
 601             felec            = _mm_macc_pd(eweps,ewtabFn,_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF));
 602             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 603
 604             /* Analytical LJ-PME */
 605             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 606             ewcljrsq         = _mm_mul_pd(ewclj2,rsq00);
 607             ewclj6           = _mm_mul_pd(ewclj2,_mm_mul_pd(ewclj2,ewclj2));
 608             exponent         = avx128fma_exp_d(ewcljrsq);
 609             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 610             poly             = _mm_mul_pd(exponent,_mm_macc_pd(_mm_mul_pd(ewcljrsq,ewcljrsq),one_half,_mm_sub_pd(one,ewcljrsq)));
 611             /* f6A = 6 * C6grid * (1 - poly) */
 612             f6A              = _mm_mul_pd(c6grid_00,_mm_sub_pd(one,poly));
 613             /* f6B = C6grid * exponent * beta^6 */
 614             f6B              = _mm_mul_pd(_mm_mul_pd(c6grid_00,one_sixth),_mm_mul_pd(exponent,ewclj6));
 615             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 616             fvdw              = _mm_mul_pd(_mm_macc_pd(_mm_msub_pd(c12_00,rinvsix,_mm_sub_pd(c6_00,f6A)),rinvsix,f6B),rinvsq00);
 617
 618             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 619
 620             fscal            = _mm_add_pd(felec,fvdw);
 621
 622             fscal            = _mm_and_pd(fscal,cutoff_mask);
 623
 624             /* Update vectorial force */
 625             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 626             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 627             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 628
 629             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 630                                                    _mm_mul_pd(dx00,fscal),
 631                                                    _mm_mul_pd(dy00,fscal),
 632                                                    _mm_mul_pd(dz00,fscal));
 633
 634             }
 635
 636             /* Inner loop uses 63 flops */
 637         }
 638
 639         if(jidx<j_index_end)
 640         {
 641
 642             jnrA             = jjnr[jidx];
 643             j_coord_offsetA  = DIM*jnrA;
 644
 645             /* load j atom coordinates */
 646             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 647                                               &jx0,&jy0,&jz0);
 648
 649             /* Calculate displacement vector */
 650             dx00             = _mm_sub_pd(ix0,jx0);
 651             dy00             = _mm_sub_pd(iy0,jy0);
 652             dz00             = _mm_sub_pd(iz0,jz0);
 653
 654             /* Calculate squared distance and things based on it */
 655             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 656
 657             rinv00           = avx128fma_invsqrt_d(rsq00);
 658
 659             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 660
 661             /* Load parameters for j particles */
 662             jq0              = _mm_load_sd(charge+jnrA+0);
 663             vdwjidx0A        = 2*vdwtype[jnrA+0];
 664
 665             /**************************
 666              * CALCULATE INTERACTIONS *
 667              **************************/
 668
 669             if (gmx_mm_any_lt(rsq00,rcutoff2))
 670             {
 671
 672             r00              = _mm_mul_pd(rsq00,rinv00);
 673
 674             /* Compute parameters for interactions between i and j atoms */
 675             qq00             = _mm_mul_pd(iq0,jq0);
 676             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 677             c6grid_00       = gmx_mm_load_1real_pd(vdwgridparam+vdwioffset0+vdwjidx0A);
 678
 679             /* EWALD ELECTROSTATICS */
 680
 681             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 682             ewrt             = _mm_mul_pd(r00,ewtabscale);
 683             ewitab           = _mm_cvttpd_epi32(ewrt);
 684 #ifdef __XOP__
 685             eweps            = _mm_frcz_pd(ewrt);
 686 #else
 687             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 688 #endif
 689             twoeweps         = _mm_add_pd(eweps,eweps);
 690             gmx_mm_load_1pair_swizzle_pd(ewtab+_mm_extract_epi32(ewitab,0),&ewtabF,&ewtabFn);
 691             felec            = _mm_macc_pd(eweps,ewtabFn,_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF));
 692             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 693
 694             /* Analytical LJ-PME */
 695             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 696             ewcljrsq         = _mm_mul_pd(ewclj2,rsq00);
 697             ewclj6           = _mm_mul_pd(ewclj2,_mm_mul_pd(ewclj2,ewclj2));
 698             exponent         = avx128fma_exp_d(ewcljrsq);
 699             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 700             poly             = _mm_mul_pd(exponent,_mm_macc_pd(_mm_mul_pd(ewcljrsq,ewcljrsq),one_half,_mm_sub_pd(one,ewcljrsq)));
 701             /* f6A = 6 * C6grid * (1 - poly) */
 702             f6A              = _mm_mul_pd(c6grid_00,_mm_sub_pd(one,poly));
 703             /* f6B = C6grid * exponent * beta^6 */
 704             f6B              = _mm_mul_pd(_mm_mul_pd(c6grid_00,one_sixth),_mm_mul_pd(exponent,ewclj6));
 705             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 706             fvdw              = _mm_mul_pd(_mm_macc_pd(_mm_msub_pd(c12_00,rinvsix,_mm_sub_pd(c6_00,f6A)),rinvsix,f6B),rinvsq00);
 707
 708             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 709
 710             fscal            = _mm_add_pd(felec,fvdw);
 711
 712             fscal            = _mm_and_pd(fscal,cutoff_mask);
 713
 714             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 715
 716             /* Update vectorial force */
 717             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 718             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 719             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 720
 721             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 722                                                    _mm_mul_pd(dx00,fscal),
 723                                                    _mm_mul_pd(dy00,fscal),
 724                                                    _mm_mul_pd(dz00,fscal));
 725
 726             }
 727
 728             /* Inner loop uses 63 flops */
 729         }
 730
 731         /* End of innermost loop */
 732
 733         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 734                                               f+i_coord_offset,fshift+i_shift_offset);
 735
 736         /* Increment number of inner iterations */
 737         inneriter                  += j_index_end - j_index_start;
 738
 739         /* Outer loop uses 7 flops */
 740     }
 741
 742     /* Increment number of outer iterations */
 743     outeriter        += nri;
 744
 745     /* Update outer/inner flops */
 746
 747     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*63);
 748 }