src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecEwSh_VdwLJEwSh_GeomP1P1_avx_128_fma_double.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_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/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_avx_128_fma_double.h"
  50 #include "kernelutil_x86_avx_128_fma_double.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJEwSh_GeomP1P1_VF_avx_128_fma_double
  54  * Electrostatics interaction: Ewald
  55  * VdW interaction:            LJEwald
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecEwSh_VdwLJEwSh_GeomP1P1_VF_avx_128_fma_double
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB;
  77     int              j_coord_offsetA,j_coord_offsetB;
  78     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  79     real             rcutoff_scalar;
  80     real             *shiftvec,*fshift,*x,*f;
  81     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwjidx0A,vdwjidx0B;
  85     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  86     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  87     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  88     real             *charge;
  89     int              nvdwtype;
  90     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  91     int              *vdwtype;
  92     real             *vdwparam;
  93     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
  94     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
  95     __m128d           c6grid_00;
  96     real             *vdwgridparam;
  97     __m128d           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
  98     __m128d           one_half  = _mm_set1_pd(0.5);
  99     __m128d           minus_one = _mm_set1_pd(-1.0);
 100     __m128i          ewitab;
 101     __m128d          ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 102     real             *ewtab;
 103     __m128d          dummy_mask,cutoff_mask;
 104     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 105     __m128d          one     = _mm_set1_pd(1.0);
 106     __m128d          two     = _mm_set1_pd(2.0);
 107     x                = xx[0];
 108     f                = ff[0];
 109
 110     nri              = nlist->nri;
 111     iinr             = nlist->iinr;
 112     jindex           = nlist->jindex;
 113     jjnr             = nlist->jjnr;
 114     shiftidx         = nlist->shift;
 115     gid              = nlist->gid;
 116     shiftvec         = fr->shift_vec[0];
 117     fshift           = fr->fshift[0];
 118     facel            = _mm_set1_pd(fr->epsfac);
 119     charge           = mdatoms->chargeA;
 120     nvdwtype         = fr->ntype;
 121     vdwparam         = fr->nbfp;
 122     vdwtype          = mdatoms->typeA;
 123     vdwgridparam     = fr->ljpme_c6grid;
 124     sh_lj_ewald      = _mm_set1_pd(fr->ic->sh_lj_ewald);
 125     ewclj            = _mm_set1_pd(fr->ewaldcoeff_lj);
 126     ewclj2           = _mm_mul_pd(minus_one,_mm_mul_pd(ewclj,ewclj));
 127
 128     sh_ewald         = _mm_set1_pd(fr->ic->sh_ewald);
 129     ewtab            = fr->ic->tabq_coul_FDV0;
 130     ewtabscale       = _mm_set1_pd(fr->ic->tabq_scale);
 131     ewtabhalfspace   = _mm_set1_pd(0.5/fr->ic->tabq_scale);
 132
 133     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 134     rcutoff_scalar   = fr->rcoulomb;
 135     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 136     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 137
 138     sh_vdw_invrcut6  = _mm_set1_pd(fr->ic->sh_invrc6);
 139     rvdw             = _mm_set1_pd(fr->rvdw);
 140
 141     /* Avoid stupid compiler warnings */
 142     jnrA = jnrB = 0;
 143     j_coord_offsetA = 0;
 144     j_coord_offsetB = 0;
 145
 146     outeriter        = 0;
 147     inneriter        = 0;
 148
 149     /* Start outer loop over neighborlists */
 150     for(iidx=0; iidx<nri; iidx++)
 151     {
 152         /* Load shift vector for this list */
 153         i_shift_offset   = DIM*shiftidx[iidx];
 154
 155         /* Load limits for loop over neighbors */
 156         j_index_start    = jindex[iidx];
 157         j_index_end      = jindex[iidx+1];
 158
 159         /* Get outer coordinate index */
 160         inr              = iinr[iidx];
 161         i_coord_offset   = DIM*inr;
 162
 163         /* Load i particle coords and add shift vector */
 164         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 165
 166         fix0             = _mm_setzero_pd();
 167         fiy0             = _mm_setzero_pd();
 168         fiz0             = _mm_setzero_pd();
 169
 170         /* Load parameters for i particles */
 171         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 172         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 173
 174         /* Reset potential sums */
 175         velecsum         = _mm_setzero_pd();
 176         vvdwsum          = _mm_setzero_pd();
 177
 178         /* Start inner kernel loop */
 179         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 180         {
 181
 182             /* Get j neighbor index, and coordinate index */
 183             jnrA             = jjnr[jidx];
 184             jnrB             = jjnr[jidx+1];
 185             j_coord_offsetA  = DIM*jnrA;
 186             j_coord_offsetB  = DIM*jnrB;
 187
 188             /* load j atom coordinates */
 189             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 190                                               &jx0,&jy0,&jz0);
 191
 192             /* Calculate displacement vector */
 193             dx00             = _mm_sub_pd(ix0,jx0);
 194             dy00             = _mm_sub_pd(iy0,jy0);
 195             dz00             = _mm_sub_pd(iz0,jz0);
 196
 197             /* Calculate squared distance and things based on it */
 198             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 199
 200             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 201
 202             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 203
 204             /* Load parameters for j particles */
 205             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 206             vdwjidx0A        = 2*vdwtype[jnrA+0];
 207             vdwjidx0B        = 2*vdwtype[jnrB+0];
 208
 209             /**************************
 210              * CALCULATE INTERACTIONS *
 211              **************************/
 212
 213             if (gmx_mm_any_lt(rsq00,rcutoff2))
 214             {
 215
 216             r00              = _mm_mul_pd(rsq00,rinv00);
 217
 218             /* Compute parameters for interactions between i and j atoms */
 219             qq00             = _mm_mul_pd(iq0,jq0);
 220             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 221                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 222             c6grid_00       = gmx_mm_load_2real_swizzle_pd(vdwgridparam+vdwioffset0+vdwjidx0A,
 223                                                                vdwgridparam+vdwioffset0+vdwjidx0B);
 224
 225             /* EWALD ELECTROSTATICS */
 226
 227             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 228             ewrt             = _mm_mul_pd(r00,ewtabscale);
 229             ewitab           = _mm_cvttpd_epi32(ewrt);
 230 #ifdef __XOP__
 231             eweps            = _mm_frcz_pd(ewrt);
 232 #else
 233             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 234 #endif
 235             twoeweps         = _mm_add_pd(eweps,eweps);
 236             ewitab           = _mm_slli_epi32(ewitab,2);
 237             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 238             ewtabD           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,1) );
 239             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 240             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 241             ewtabFn          = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,1) +2);
 242             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 243             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 244             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 245             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_sub_pd(rinv00,sh_ewald),velec));
 246             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 247
 248             /* Analytical LJ-PME */
 249             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 250             ewcljrsq         = _mm_mul_pd(ewclj2,rsq00);
 251             ewclj6           = _mm_mul_pd(ewclj2,_mm_mul_pd(ewclj2,ewclj2));
 252             exponent         = gmx_simd_exp_d(ewcljrsq);
 253             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 254             poly             = _mm_mul_pd(exponent,_mm_macc_pd(_mm_mul_pd(ewcljrsq,ewcljrsq),one_half,_mm_sub_pd(one,ewcljrsq)));
 255             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 256             vvdw6            = _mm_mul_pd(_mm_macc_pd(-c6grid_00,_mm_sub_pd(one,poly),c6_00),rinvsix);
 257             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 258             vvdw             = _mm_msub_pd(_mm_nmacc_pd(c12_00,_mm_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 259                                _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));
 260             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 261             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);
 262
 263             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 264
 265             /* Update potential sum for this i atom from the interaction with this j atom. */
 266             velec            = _mm_and_pd(velec,cutoff_mask);
 267             velecsum         = _mm_add_pd(velecsum,velec);
 268             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 269             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 270
 271             fscal            = _mm_add_pd(felec,fvdw);
 272
 273             fscal            = _mm_and_pd(fscal,cutoff_mask);
 274
 275             /* Update vectorial force */
 276             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 277             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 278             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 279
 280             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 281                                                    _mm_mul_pd(dx00,fscal),
 282                                                    _mm_mul_pd(dy00,fscal),
 283                                                    _mm_mul_pd(dz00,fscal));
 284
 285             }
 286
 287             /* Inner loop uses 78 flops */
 288         }
 289
 290         if(jidx<j_index_end)
 291         {
 292
 293             jnrA             = jjnr[jidx];
 294             j_coord_offsetA  = DIM*jnrA;
 295
 296             /* load j atom coordinates */
 297             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 298                                               &jx0,&jy0,&jz0);
 299
 300             /* Calculate displacement vector */
 301             dx00             = _mm_sub_pd(ix0,jx0);
 302             dy00             = _mm_sub_pd(iy0,jy0);
 303             dz00             = _mm_sub_pd(iz0,jz0);
 304
 305             /* Calculate squared distance and things based on it */
 306             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 307
 308             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 309
 310             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 311
 312             /* Load parameters for j particles */
 313             jq0              = _mm_load_sd(charge+jnrA+0);
 314             vdwjidx0A        = 2*vdwtype[jnrA+0];
 315
 316             /**************************
 317              * CALCULATE INTERACTIONS *
 318              **************************/
 319
 320             if (gmx_mm_any_lt(rsq00,rcutoff2))
 321             {
 322
 323             r00              = _mm_mul_pd(rsq00,rinv00);
 324
 325             /* Compute parameters for interactions between i and j atoms */
 326             qq00             = _mm_mul_pd(iq0,jq0);
 327             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 328             c6grid_00       = gmx_mm_load_1real_pd(vdwgridparam+vdwioffset0+vdwjidx0A);
 329
 330             /* EWALD ELECTROSTATICS */
 331
 332             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 333             ewrt             = _mm_mul_pd(r00,ewtabscale);
 334             ewitab           = _mm_cvttpd_epi32(ewrt);
 335 #ifdef __XOP__
 336             eweps            = _mm_frcz_pd(ewrt);
 337 #else
 338             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 339 #endif
 340             twoeweps         = _mm_add_pd(eweps,eweps);
 341             ewitab           = _mm_slli_epi32(ewitab,2);
 342             ewtabF           = _mm_load_pd( ewtab + _mm_extract_epi32(ewitab,0) );
 343             ewtabD           = _mm_setzero_pd();
 344             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 345             ewtabV           = _mm_load_sd( ewtab + _mm_extract_epi32(ewitab,0) +2);
 346             ewtabFn          = _mm_setzero_pd();
 347             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 348             felec            = _mm_macc_pd(eweps,ewtabD,ewtabF);
 349             velec            = _mm_nmacc_pd(_mm_mul_pd(ewtabhalfspace,eweps) ,_mm_add_pd(ewtabF,felec), ewtabV);
 350             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_sub_pd(rinv00,sh_ewald),velec));
 351             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 352
 353             /* Analytical LJ-PME */
 354             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 355             ewcljrsq         = _mm_mul_pd(ewclj2,rsq00);
 356             ewclj6           = _mm_mul_pd(ewclj2,_mm_mul_pd(ewclj2,ewclj2));
 357             exponent         = gmx_simd_exp_d(ewcljrsq);
 358             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 359             poly             = _mm_mul_pd(exponent,_mm_macc_pd(_mm_mul_pd(ewcljrsq,ewcljrsq),one_half,_mm_sub_pd(one,ewcljrsq)));
 360             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 361             vvdw6            = _mm_mul_pd(_mm_macc_pd(-c6grid_00,_mm_sub_pd(one,poly),c6_00),rinvsix);
 362             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 363             vvdw             = _mm_msub_pd(_mm_nmacc_pd(c12_00,_mm_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 364                                _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));
 365             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 366             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);
 367
 368             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 369
 370             /* Update potential sum for this i atom from the interaction with this j atom. */
 371             velec            = _mm_and_pd(velec,cutoff_mask);
 372             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 373             velecsum         = _mm_add_pd(velecsum,velec);
 374             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 375             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 376             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 377
 378             fscal            = _mm_add_pd(felec,fvdw);
 379
 380             fscal            = _mm_and_pd(fscal,cutoff_mask);
 381
 382             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 383
 384             /* Update vectorial force */
 385             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 386             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 387             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 388
 389             gmx_mm_decrement_1rvec_1ptr_swizzle_pd(f+j_coord_offsetA,
 390                                                    _mm_mul_pd(dx00,fscal),
 391                                                    _mm_mul_pd(dy00,fscal),
 392                                                    _mm_mul_pd(dz00,fscal));
 393
 394             }
 395
 396             /* Inner loop uses 78 flops */
 397         }
 398
 399         /* End of innermost loop */
 400
 401         gmx_mm_update_iforce_1atom_swizzle_pd(fix0,fiy0,fiz0,
 402                                               f+i_coord_offset,fshift+i_shift_offset);
 403
 404         ggid                        = gid[iidx];
 405         /* Update potential energies */
 406         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 407         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 408
 409         /* Increment number of inner iterations */
 410         inneriter                  += j_index_end - j_index_start;
 411
 412         /* Outer loop uses 9 flops */
 413     }
 414
 415     /* Increment number of outer iterations */
 416     outeriter        += nri;
 417
 418     /* Update outer/inner flops */
 419
 420     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*78);
 421 }
 422 /*
 423  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJEwSh_GeomP1P1_F_avx_128_fma_double
 424  * Electrostatics interaction: Ewald
 425  * VdW interaction:            LJEwald
 426  * Geometry:                   Particle-Particle
 427  * Calculate force/pot:        Force
 428  */
 429 void
 430 nb_kernel_ElecEwSh_VdwLJEwSh_GeomP1P1_F_avx_128_fma_double
 431                     (t_nblist                    * gmx_restrict       nlist,
 432                      rvec                        * gmx_restrict          xx,
 433                      rvec                        * gmx_restrict          ff,
 434                      t_forcerec                  * gmx_restrict          fr,
 435                      t_mdatoms                   * gmx_restrict     mdatoms,
 436                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 437                      t_nrnb                      * gmx_restrict        nrnb)
 438 {
 439     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 440      * just 0 for non-waters.
 441      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 442      * jnr indices corresponding to data put in the four positions in the SIMD register.
 443      */
 444     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 445     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 446     int              jnrA,jnrB;
 447     int              j_coord_offsetA,j_coord_offsetB;
 448     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 449     real             rcutoff_scalar;
 450     real             *shiftvec,*fshift,*x,*f;
 451     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 452     int              vdwioffset0;
 453     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 454     int              vdwjidx0A,vdwjidx0B;
 455     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 456     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 457     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 458     real             *charge;
 459     int              nvdwtype;
 460     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 461     int              *vdwtype;
 462     real             *vdwparam;
 463     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 464     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 465     __m128d           c6grid_00;
 466     real             *vdwgridparam;
 467     __m128d           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 468     __m128d           one_half  = _mm_set1_pd(0.5);
 469     __m128d           minus_one = _mm_set1_pd(-1.0);
 470     __m128i          ewitab;
 471     __m128d          ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 472     real             *ewtab;
 473     __m128d          dummy_mask,cutoff_mask;
 474     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 475     __m128d          one     = _mm_set1_pd(1.0);
 476     __m128d          two     = _mm_set1_pd(2.0);
 477     x                = xx[0];
 478     f                = ff[0];
 479
 480     nri              = nlist->nri;
 481     iinr             = nlist->iinr;
 482     jindex           = nlist->jindex;
 483     jjnr             = nlist->jjnr;
 484     shiftidx         = nlist->shift;
 485     gid              = nlist->gid;
 486     shiftvec         = fr->shift_vec[0];
 487     fshift           = fr->fshift[0];
 488     facel            = _mm_set1_pd(fr->epsfac);
 489     charge           = mdatoms->chargeA;
 490     nvdwtype         = fr->ntype;
 491     vdwparam         = fr->nbfp;
 492     vdwtype          = mdatoms->typeA;
 493     vdwgridparam     = fr->ljpme_c6grid;
 494     sh_lj_ewald      = _mm_set1_pd(fr->ic->sh_lj_ewald);
 495     ewclj            = _mm_set1_pd(fr->ewaldcoeff_lj);
 496     ewclj2           = _mm_mul_pd(minus_one,_mm_mul_pd(ewclj,ewclj));
 497
 498     sh_ewald         = _mm_set1_pd(fr->ic->sh_ewald);
 499     ewtab            = fr->ic->tabq_coul_F;
 500     ewtabscale       = _mm_set1_pd(fr->ic->tabq_scale);
 501     ewtabhalfspace   = _mm_set1_pd(0.5/fr->ic->tabq_scale);
 502
 503     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 504     rcutoff_scalar   = fr->rcoulomb;
 505     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 506     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 507
 508     sh_vdw_invrcut6  = _mm_set1_pd(fr->ic->sh_invrc6);
 509     rvdw             = _mm_set1_pd(fr->rvdw);
 510
 511     /* Avoid stupid compiler warnings */
 512     jnrA = jnrB = 0;
 513     j_coord_offsetA = 0;
 514     j_coord_offsetB = 0;
 515
 516     outeriter        = 0;
 517     inneriter        = 0;
 518
 519     /* Start outer loop over neighborlists */
 520     for(iidx=0; iidx<nri; iidx++)
 521     {
 522         /* Load shift vector for this list */
 523         i_shift_offset   = DIM*shiftidx[iidx];
 524
 525         /* Load limits for loop over neighbors */
 526         j_index_start    = jindex[iidx];
 527         j_index_end      = jindex[iidx+1];
 528
 529         /* Get outer coordinate index */
 530         inr              = iinr[iidx];
 531         i_coord_offset   = DIM*inr;
 532
 533         /* Load i particle coords and add shift vector */
 534         gmx_mm_load_shift_and_1rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 535
 536         fix0             = _mm_setzero_pd();
 537         fiy0             = _mm_setzero_pd();
 538         fiz0             = _mm_setzero_pd();
 539
 540         /* Load parameters for i particles */
 541         iq0              = _mm_mul_pd(facel,_mm_load1_pd(charge+inr+0));
 542         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 543
 544         /* Start inner kernel loop */
 545         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 546         {
 547
 548             /* Get j neighbor index, and coordinate index */
 549             jnrA             = jjnr[jidx];
 550             jnrB             = jjnr[jidx+1];
 551             j_coord_offsetA  = DIM*jnrA;
 552             j_coord_offsetB  = DIM*jnrB;
 553
 554             /* load j atom coordinates */
 555             gmx_mm_load_1rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 556                                               &jx0,&jy0,&jz0);
 557
 558             /* Calculate displacement vector */
 559             dx00             = _mm_sub_pd(ix0,jx0);
 560             dy00             = _mm_sub_pd(iy0,jy0);
 561             dz00             = _mm_sub_pd(iz0,jz0);
 562
 563             /* Calculate squared distance and things based on it */
 564             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 565
 566             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 567
 568             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 569
 570             /* Load parameters for j particles */
 571             jq0              = gmx_mm_load_2real_swizzle_pd(charge+jnrA+0,charge+jnrB+0);
 572             vdwjidx0A        = 2*vdwtype[jnrA+0];
 573             vdwjidx0B        = 2*vdwtype[jnrB+0];
 574
 575             /**************************
 576              * CALCULATE INTERACTIONS *
 577              **************************/
 578
 579             if (gmx_mm_any_lt(rsq00,rcutoff2))
 580             {
 581
 582             r00              = _mm_mul_pd(rsq00,rinv00);
 583
 584             /* Compute parameters for interactions between i and j atoms */
 585             qq00             = _mm_mul_pd(iq0,jq0);
 586             gmx_mm_load_2pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,
 587                                          vdwparam+vdwioffset0+vdwjidx0B,&c6_00,&c12_00);
 588             c6grid_00       = gmx_mm_load_2real_swizzle_pd(vdwgridparam+vdwioffset0+vdwjidx0A,
 589                                                                vdwgridparam+vdwioffset0+vdwjidx0B);
 590
 591             /* EWALD ELECTROSTATICS */
 592
 593             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 594             ewrt             = _mm_mul_pd(r00,ewtabscale);
 595             ewitab           = _mm_cvttpd_epi32(ewrt);
 596 #ifdef __XOP__
 597             eweps            = _mm_frcz_pd(ewrt);
 598 #else
 599             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 600 #endif
 601             twoeweps         = _mm_add_pd(eweps,eweps);
 602             gmx_mm_load_2pair_swizzle_pd(ewtab+_mm_extract_epi32(ewitab,0),ewtab+_mm_extract_epi32(ewitab,1),
 603                                          &ewtabF,&ewtabFn);
 604             felec            = _mm_macc_pd(eweps,ewtabFn,_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF));
 605             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 606
 607             /* Analytical LJ-PME */
 608             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 609             ewcljrsq         = _mm_mul_pd(ewclj2,rsq00);
 610             ewclj6           = _mm_mul_pd(ewclj2,_mm_mul_pd(ewclj2,ewclj2));
 611             exponent         = gmx_simd_exp_d(ewcljrsq);
 612             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 613             poly             = _mm_mul_pd(exponent,_mm_macc_pd(_mm_mul_pd(ewcljrsq,ewcljrsq),one_half,_mm_sub_pd(one,ewcljrsq)));
 614             /* f6A = 6 * C6grid * (1 - poly) */
 615             f6A              = _mm_mul_pd(c6grid_00,_mm_sub_pd(one,poly));
 616             /* f6B = C6grid * exponent * beta^6 */
 617             f6B              = _mm_mul_pd(_mm_mul_pd(c6grid_00,one_sixth),_mm_mul_pd(exponent,ewclj6));
 618             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 619             fvdw              = _mm_mul_pd(_mm_macc_pd(_mm_msub_pd(c12_00,rinvsix,_mm_sub_pd(c6_00,f6A)),rinvsix,f6B),rinvsq00);
 620
 621             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 622
 623             fscal            = _mm_add_pd(felec,fvdw);
 624
 625             fscal            = _mm_and_pd(fscal,cutoff_mask);
 626
 627             /* Update vectorial force */
 628             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 629             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 630             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 631
 632             gmx_mm_decrement_1rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,
 633                                                    _mm_mul_pd(dx00,fscal),
 634                                                    _mm_mul_pd(dy00,fscal),
 635                                                    _mm_mul_pd(dz00,fscal));
 636
 637             }
 638
 639             /* Inner loop uses 63 flops */
 640         }
 641
 642         if(jidx<j_index_end)
 643         {
 644
 645             jnrA             = jjnr[jidx];
 646             j_coord_offsetA  = DIM*jnrA;
 647
 648             /* load j atom coordinates */
 649             gmx_mm_load_1rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 650                                               &jx0,&jy0,&jz0);
 651
 652             /* Calculate displacement vector */
 653             dx00             = _mm_sub_pd(ix0,jx0);
 654             dy00             = _mm_sub_pd(iy0,jy0);
 655             dz00             = _mm_sub_pd(iz0,jz0);
 656
 657             /* Calculate squared distance and things based on it */
 658             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 659
 660             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 661
 662             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 663
 664             /* Load parameters for j particles */
 665             jq0              = _mm_load_sd(charge+jnrA+0);
 666             vdwjidx0A        = 2*vdwtype[jnrA+0];
 667
 668             /**************************
 669              * CALCULATE INTERACTIONS *
 670              **************************/
 671
 672             if (gmx_mm_any_lt(rsq00,rcutoff2))
 673             {
 674
 675             r00              = _mm_mul_pd(rsq00,rinv00);
 676
 677             /* Compute parameters for interactions between i and j atoms */
 678             qq00             = _mm_mul_pd(iq0,jq0);
 679             gmx_mm_load_1pair_swizzle_pd(vdwparam+vdwioffset0+vdwjidx0A,&c6_00,&c12_00);
 680             c6grid_00       = gmx_mm_load_1real_pd(vdwgridparam+vdwioffset0+vdwjidx0A);
 681
 682             /* EWALD ELECTROSTATICS */
 683
 684             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 685             ewrt             = _mm_mul_pd(r00,ewtabscale);
 686             ewitab           = _mm_cvttpd_epi32(ewrt);
 687 #ifdef __XOP__
 688             eweps            = _mm_frcz_pd(ewrt);
 689 #else
 690             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 691 #endif
 692             twoeweps         = _mm_add_pd(eweps,eweps);
 693             gmx_mm_load_1pair_swizzle_pd(ewtab+_mm_extract_epi32(ewitab,0),&ewtabF,&ewtabFn);
 694             felec            = _mm_macc_pd(eweps,ewtabFn,_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF));
 695             felec            = _mm_mul_pd(_mm_mul_pd(qq00,rinv00),_mm_sub_pd(rinvsq00,felec));
 696
 697             /* Analytical LJ-PME */
 698             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 699             ewcljrsq         = _mm_mul_pd(ewclj2,rsq00);
 700             ewclj6           = _mm_mul_pd(ewclj2,_mm_mul_pd(ewclj2,ewclj2));
 701             exponent         = gmx_simd_exp_d(ewcljrsq);
 702             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 703             poly             = _mm_mul_pd(exponent,_mm_macc_pd(_mm_mul_pd(ewcljrsq,ewcljrsq),one_half,_mm_sub_pd(one,ewcljrsq)));
 704             /* f6A = 6 * C6grid * (1 - poly) */
 705             f6A              = _mm_mul_pd(c6grid_00,_mm_sub_pd(one,poly));
 706             /* f6B = C6grid * exponent * beta^6 */
 707             f6B              = _mm_mul_pd(_mm_mul_pd(c6grid_00,one_sixth),_mm_mul_pd(exponent,ewclj6));
 708             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 709             fvdw              = _mm_mul_pd(_mm_macc_pd(_mm_msub_pd(c12_00,rinvsix,_mm_sub_pd(c6_00,f6A)),rinvsix,f6B),rinvsq00);
 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 63 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*63);
 751 }