src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEwSh_VdwLJSh_GeomP1P1_avx_128_fma_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_single kernel generator.
  37  */
  38 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  50 #include "kernelutil_x86_avx_128_fma_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJSh_GeomP1P1_VF_avx_128_fma_single
  54  * Electrostatics interaction: Ewald
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecEwSh_VdwLJSh_GeomP1P1_VF_avx_128_fma_single
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  83     real             scratch[4*DIM];
  84     __m128           fscal,rcutoff,rcutoff2,jidxall;
  85     int              vdwioffset0;
  86     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  88     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  90     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  91     real             *charge;
  92     int              nvdwtype;
  93     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  94     int              *vdwtype;
  95     real             *vdwparam;
  96     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  97     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  98     __m128i          ewitab;
  99     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 100     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 101     real             *ewtab;
 102     __m128           dummy_mask,cutoff_mask;
 103     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 104     __m128           one     = _mm_set1_ps(1.0);
 105     __m128           two     = _mm_set1_ps(2.0);
 106     x                = xx[0];
 107     f                = ff[0];
 108
 109     nri              = nlist->nri;
 110     iinr             = nlist->iinr;
 111     jindex           = nlist->jindex;
 112     jjnr             = nlist->jjnr;
 113     shiftidx         = nlist->shift;
 114     gid              = nlist->gid;
 115     shiftvec         = fr->shift_vec[0];
 116     fshift           = fr->fshift[0];
 117     facel            = _mm_set1_ps(fr->epsfac);
 118     charge           = mdatoms->chargeA;
 119     nvdwtype         = fr->ntype;
 120     vdwparam         = fr->nbfp;
 121     vdwtype          = mdatoms->typeA;
 122
 123     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 124     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 125     beta2            = _mm_mul_ps(beta,beta);
 126     beta3            = _mm_mul_ps(beta,beta2);
 127     ewtab            = fr->ic->tabq_coul_FDV0;
 128     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 129     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 130
 131     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 132     rcutoff_scalar   = fr->rcoulomb;
 133     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 134     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 135
 136     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 137     rvdw             = _mm_set1_ps(fr->rvdw);
 138
 139     /* Avoid stupid compiler warnings */
 140     jnrA = jnrB = jnrC = jnrD = 0;
 141     j_coord_offsetA = 0;
 142     j_coord_offsetB = 0;
 143     j_coord_offsetC = 0;
 144     j_coord_offsetD = 0;
 145
 146     outeriter        = 0;
 147     inneriter        = 0;
 148
 149     for(iidx=0;iidx<4*DIM;iidx++)
 150     {
 151         scratch[iidx] = 0.0;
 152     }
 153
 154     /* Start outer loop over neighborlists */
 155     for(iidx=0; iidx<nri; iidx++)
 156     {
 157         /* Load shift vector for this list */
 158         i_shift_offset   = DIM*shiftidx[iidx];
 159
 160         /* Load limits for loop over neighbors */
 161         j_index_start    = jindex[iidx];
 162         j_index_end      = jindex[iidx+1];
 163
 164         /* Get outer coordinate index */
 165         inr              = iinr[iidx];
 166         i_coord_offset   = DIM*inr;
 167
 168         /* Load i particle coords and add shift vector */
 169         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 170
 171         fix0             = _mm_setzero_ps();
 172         fiy0             = _mm_setzero_ps();
 173         fiz0             = _mm_setzero_ps();
 174
 175         /* Load parameters for i particles */
 176         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 177         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 178
 179         /* Reset potential sums */
 180         velecsum         = _mm_setzero_ps();
 181         vvdwsum          = _mm_setzero_ps();
 182
 183         /* Start inner kernel loop */
 184         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 185         {
 186
 187             /* Get j neighbor index, and coordinate index */
 188             jnrA             = jjnr[jidx];
 189             jnrB             = jjnr[jidx+1];
 190             jnrC             = jjnr[jidx+2];
 191             jnrD             = jjnr[jidx+3];
 192             j_coord_offsetA  = DIM*jnrA;
 193             j_coord_offsetB  = DIM*jnrB;
 194             j_coord_offsetC  = DIM*jnrC;
 195             j_coord_offsetD  = DIM*jnrD;
 196
 197             /* load j atom coordinates */
 198             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 199                                               x+j_coord_offsetC,x+j_coord_offsetD,
 200                                               &jx0,&jy0,&jz0);
 201
 202             /* Calculate displacement vector */
 203             dx00             = _mm_sub_ps(ix0,jx0);
 204             dy00             = _mm_sub_ps(iy0,jy0);
 205             dz00             = _mm_sub_ps(iz0,jz0);
 206
 207             /* Calculate squared distance and things based on it */
 208             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 209
 210             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 211
 212             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 213
 214             /* Load parameters for j particles */
 215             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 216                                                               charge+jnrC+0,charge+jnrD+0);
 217             vdwjidx0A        = 2*vdwtype[jnrA+0];
 218             vdwjidx0B        = 2*vdwtype[jnrB+0];
 219             vdwjidx0C        = 2*vdwtype[jnrC+0];
 220             vdwjidx0D        = 2*vdwtype[jnrD+0];
 221
 222             /**************************
 223              * CALCULATE INTERACTIONS *
 224              **************************/
 225
 226             if (gmx_mm_any_lt(rsq00,rcutoff2))
 227             {
 228
 229             r00              = _mm_mul_ps(rsq00,rinv00);
 230
 231             /* Compute parameters for interactions between i and j atoms */
 232             qq00             = _mm_mul_ps(iq0,jq0);
 233             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 234                                          vdwparam+vdwioffset0+vdwjidx0B,
 235                                          vdwparam+vdwioffset0+vdwjidx0C,
 236                                          vdwparam+vdwioffset0+vdwjidx0D,
 237                                          &c6_00,&c12_00);
 238
 239             /* EWALD ELECTROSTATICS */
 240
 241             /* Analytical PME correction */
 242             zeta2            = _mm_mul_ps(beta2,rsq00);
 243             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 244             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 245             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 246             felec            = _mm_mul_ps(qq00,felec);
 247             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 248             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv00,sh_ewald));
 249             velec            = _mm_mul_ps(qq00,velec);
 250
 251             /* LENNARD-JONES DISPERSION/REPULSION */
 252
 253             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 254             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 255             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 256             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 257                                           _mm_mul_ps( _mm_nmacc_ps(c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 258             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 259
 260             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 261
 262             /* Update potential sum for this i atom from the interaction with this j atom. */
 263             velec            = _mm_and_ps(velec,cutoff_mask);
 264             velecsum         = _mm_add_ps(velecsum,velec);
 265             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 266             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 267
 268             fscal            = _mm_add_ps(felec,fvdw);
 269
 270             fscal            = _mm_and_ps(fscal,cutoff_mask);
 271
 272              /* Update vectorial force */
 273             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 274             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 275             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 276
 277             fjptrA             = f+j_coord_offsetA;
 278             fjptrB             = f+j_coord_offsetB;
 279             fjptrC             = f+j_coord_offsetC;
 280             fjptrD             = f+j_coord_offsetD;
 281             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 282                                                    _mm_mul_ps(dx00,fscal),
 283                                                    _mm_mul_ps(dy00,fscal),
 284                                                    _mm_mul_ps(dz00,fscal));
 285
 286             }
 287
 288             /* Inner loop uses 51 flops */
 289         }
 290
 291         if(jidx<j_index_end)
 292         {
 293
 294             /* Get j neighbor index, and coordinate index */
 295             jnrlistA         = jjnr[jidx];
 296             jnrlistB         = jjnr[jidx+1];
 297             jnrlistC         = jjnr[jidx+2];
 298             jnrlistD         = jjnr[jidx+3];
 299             /* Sign of each element will be negative for non-real atoms.
 300              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 301              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 302              */
 303             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 304             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 305             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 306             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 307             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 308             j_coord_offsetA  = DIM*jnrA;
 309             j_coord_offsetB  = DIM*jnrB;
 310             j_coord_offsetC  = DIM*jnrC;
 311             j_coord_offsetD  = DIM*jnrD;
 312
 313             /* load j atom coordinates */
 314             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 315                                               x+j_coord_offsetC,x+j_coord_offsetD,
 316                                               &jx0,&jy0,&jz0);
 317
 318             /* Calculate displacement vector */
 319             dx00             = _mm_sub_ps(ix0,jx0);
 320             dy00             = _mm_sub_ps(iy0,jy0);
 321             dz00             = _mm_sub_ps(iz0,jz0);
 322
 323             /* Calculate squared distance and things based on it */
 324             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 325
 326             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 327
 328             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 329
 330             /* Load parameters for j particles */
 331             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 332                                                               charge+jnrC+0,charge+jnrD+0);
 333             vdwjidx0A        = 2*vdwtype[jnrA+0];
 334             vdwjidx0B        = 2*vdwtype[jnrB+0];
 335             vdwjidx0C        = 2*vdwtype[jnrC+0];
 336             vdwjidx0D        = 2*vdwtype[jnrD+0];
 337
 338             /**************************
 339              * CALCULATE INTERACTIONS *
 340              **************************/
 341
 342             if (gmx_mm_any_lt(rsq00,rcutoff2))
 343             {
 344
 345             r00              = _mm_mul_ps(rsq00,rinv00);
 346             r00              = _mm_andnot_ps(dummy_mask,r00);
 347
 348             /* Compute parameters for interactions between i and j atoms */
 349             qq00             = _mm_mul_ps(iq0,jq0);
 350             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 351                                          vdwparam+vdwioffset0+vdwjidx0B,
 352                                          vdwparam+vdwioffset0+vdwjidx0C,
 353                                          vdwparam+vdwioffset0+vdwjidx0D,
 354                                          &c6_00,&c12_00);
 355
 356             /* EWALD ELECTROSTATICS */
 357
 358             /* Analytical PME correction */
 359             zeta2            = _mm_mul_ps(beta2,rsq00);
 360             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 361             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 362             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 363             felec            = _mm_mul_ps(qq00,felec);
 364             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 365             velec            = _mm_nmacc_ps(pmecorrV,beta,_mm_sub_ps(rinv00,sh_ewald));
 366             velec            = _mm_mul_ps(qq00,velec);
 367
 368             /* LENNARD-JONES DISPERSION/REPULSION */
 369
 370             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 371             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 372             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 373             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 374                                           _mm_mul_ps( _mm_nmacc_ps(c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 375             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 376
 377             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 378
 379             /* Update potential sum for this i atom from the interaction with this j atom. */
 380             velec            = _mm_and_ps(velec,cutoff_mask);
 381             velec            = _mm_andnot_ps(dummy_mask,velec);
 382             velecsum         = _mm_add_ps(velecsum,velec);
 383             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 384             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 385             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 386
 387             fscal            = _mm_add_ps(felec,fvdw);
 388
 389             fscal            = _mm_and_ps(fscal,cutoff_mask);
 390
 391             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 392
 393              /* Update vectorial force */
 394             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 395             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 396             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 397
 398             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 399             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 400             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 401             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 402             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 403                                                    _mm_mul_ps(dx00,fscal),
 404                                                    _mm_mul_ps(dy00,fscal),
 405                                                    _mm_mul_ps(dz00,fscal));
 406
 407             }
 408
 409             /* Inner loop uses 52 flops */
 410         }
 411
 412         /* End of innermost loop */
 413
 414         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 415                                               f+i_coord_offset,fshift+i_shift_offset);
 416
 417         ggid                        = gid[iidx];
 418         /* Update potential energies */
 419         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 420         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 421
 422         /* Increment number of inner iterations */
 423         inneriter                  += j_index_end - j_index_start;
 424
 425         /* Outer loop uses 9 flops */
 426     }
 427
 428     /* Increment number of outer iterations */
 429     outeriter        += nri;
 430
 431     /* Update outer/inner flops */
 432
 433     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*52);
 434 }
 435 /*
 436  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJSh_GeomP1P1_F_avx_128_fma_single
 437  * Electrostatics interaction: Ewald
 438  * VdW interaction:            LennardJones
 439  * Geometry:                   Particle-Particle
 440  * Calculate force/pot:        Force
 441  */
 442 void
 443 nb_kernel_ElecEwSh_VdwLJSh_GeomP1P1_F_avx_128_fma_single
 444                     (t_nblist                    * gmx_restrict       nlist,
 445                      rvec                        * gmx_restrict          xx,
 446                      rvec                        * gmx_restrict          ff,
 447                      t_forcerec                  * gmx_restrict          fr,
 448                      t_mdatoms                   * gmx_restrict     mdatoms,
 449                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 450                      t_nrnb                      * gmx_restrict        nrnb)
 451 {
 452     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 453      * just 0 for non-waters.
 454      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 455      * jnr indices corresponding to data put in the four positions in the SIMD register.
 456      */
 457     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 458     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 459     int              jnrA,jnrB,jnrC,jnrD;
 460     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 461     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 462     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 463     real             rcutoff_scalar;
 464     real             *shiftvec,*fshift,*x,*f;
 465     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 466     real             scratch[4*DIM];
 467     __m128           fscal,rcutoff,rcutoff2,jidxall;
 468     int              vdwioffset0;
 469     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 470     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 471     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 472     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 473     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 474     real             *charge;
 475     int              nvdwtype;
 476     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 477     int              *vdwtype;
 478     real             *vdwparam;
 479     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 480     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 481     __m128i          ewitab;
 482     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 483     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 484     real             *ewtab;
 485     __m128           dummy_mask,cutoff_mask;
 486     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 487     __m128           one     = _mm_set1_ps(1.0);
 488     __m128           two     = _mm_set1_ps(2.0);
 489     x                = xx[0];
 490     f                = ff[0];
 491
 492     nri              = nlist->nri;
 493     iinr             = nlist->iinr;
 494     jindex           = nlist->jindex;
 495     jjnr             = nlist->jjnr;
 496     shiftidx         = nlist->shift;
 497     gid              = nlist->gid;
 498     shiftvec         = fr->shift_vec[0];
 499     fshift           = fr->fshift[0];
 500     facel            = _mm_set1_ps(fr->epsfac);
 501     charge           = mdatoms->chargeA;
 502     nvdwtype         = fr->ntype;
 503     vdwparam         = fr->nbfp;
 504     vdwtype          = mdatoms->typeA;
 505
 506     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 507     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 508     beta2            = _mm_mul_ps(beta,beta);
 509     beta3            = _mm_mul_ps(beta,beta2);
 510     ewtab            = fr->ic->tabq_coul_F;
 511     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 512     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 513
 514     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 515     rcutoff_scalar   = fr->rcoulomb;
 516     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 517     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 518
 519     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 520     rvdw             = _mm_set1_ps(fr->rvdw);
 521
 522     /* Avoid stupid compiler warnings */
 523     jnrA = jnrB = jnrC = jnrD = 0;
 524     j_coord_offsetA = 0;
 525     j_coord_offsetB = 0;
 526     j_coord_offsetC = 0;
 527     j_coord_offsetD = 0;
 528
 529     outeriter        = 0;
 530     inneriter        = 0;
 531
 532     for(iidx=0;iidx<4*DIM;iidx++)
 533     {
 534         scratch[iidx] = 0.0;
 535     }
 536
 537     /* Start outer loop over neighborlists */
 538     for(iidx=0; iidx<nri; iidx++)
 539     {
 540         /* Load shift vector for this list */
 541         i_shift_offset   = DIM*shiftidx[iidx];
 542
 543         /* Load limits for loop over neighbors */
 544         j_index_start    = jindex[iidx];
 545         j_index_end      = jindex[iidx+1];
 546
 547         /* Get outer coordinate index */
 548         inr              = iinr[iidx];
 549         i_coord_offset   = DIM*inr;
 550
 551         /* Load i particle coords and add shift vector */
 552         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 553
 554         fix0             = _mm_setzero_ps();
 555         fiy0             = _mm_setzero_ps();
 556         fiz0             = _mm_setzero_ps();
 557
 558         /* Load parameters for i particles */
 559         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 560         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 561
 562         /* Start inner kernel loop */
 563         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 564         {
 565
 566             /* Get j neighbor index, and coordinate index */
 567             jnrA             = jjnr[jidx];
 568             jnrB             = jjnr[jidx+1];
 569             jnrC             = jjnr[jidx+2];
 570             jnrD             = jjnr[jidx+3];
 571             j_coord_offsetA  = DIM*jnrA;
 572             j_coord_offsetB  = DIM*jnrB;
 573             j_coord_offsetC  = DIM*jnrC;
 574             j_coord_offsetD  = DIM*jnrD;
 575
 576             /* load j atom coordinates */
 577             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 578                                               x+j_coord_offsetC,x+j_coord_offsetD,
 579                                               &jx0,&jy0,&jz0);
 580
 581             /* Calculate displacement vector */
 582             dx00             = _mm_sub_ps(ix0,jx0);
 583             dy00             = _mm_sub_ps(iy0,jy0);
 584             dz00             = _mm_sub_ps(iz0,jz0);
 585
 586             /* Calculate squared distance and things based on it */
 587             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 588
 589             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 590
 591             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 592
 593             /* Load parameters for j particles */
 594             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 595                                                               charge+jnrC+0,charge+jnrD+0);
 596             vdwjidx0A        = 2*vdwtype[jnrA+0];
 597             vdwjidx0B        = 2*vdwtype[jnrB+0];
 598             vdwjidx0C        = 2*vdwtype[jnrC+0];
 599             vdwjidx0D        = 2*vdwtype[jnrD+0];
 600
 601             /**************************
 602              * CALCULATE INTERACTIONS *
 603              **************************/
 604
 605             if (gmx_mm_any_lt(rsq00,rcutoff2))
 606             {
 607
 608             r00              = _mm_mul_ps(rsq00,rinv00);
 609
 610             /* Compute parameters for interactions between i and j atoms */
 611             qq00             = _mm_mul_ps(iq0,jq0);
 612             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 613                                          vdwparam+vdwioffset0+vdwjidx0B,
 614                                          vdwparam+vdwioffset0+vdwjidx0C,
 615                                          vdwparam+vdwioffset0+vdwjidx0D,
 616                                          &c6_00,&c12_00);
 617
 618             /* EWALD ELECTROSTATICS */
 619
 620             /* Analytical PME correction */
 621             zeta2            = _mm_mul_ps(beta2,rsq00);
 622             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 623             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 624             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 625             felec            = _mm_mul_ps(qq00,felec);
 626
 627             /* LENNARD-JONES DISPERSION/REPULSION */
 628
 629             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 630             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 631
 632             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 633
 634             fscal            = _mm_add_ps(felec,fvdw);
 635
 636             fscal            = _mm_and_ps(fscal,cutoff_mask);
 637
 638              /* Update vectorial force */
 639             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 640             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 641             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 642
 643             fjptrA             = f+j_coord_offsetA;
 644             fjptrB             = f+j_coord_offsetB;
 645             fjptrC             = f+j_coord_offsetC;
 646             fjptrD             = f+j_coord_offsetD;
 647             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 648                                                    _mm_mul_ps(dx00,fscal),
 649                                                    _mm_mul_ps(dy00,fscal),
 650                                                    _mm_mul_ps(dz00,fscal));
 651
 652             }
 653
 654             /* Inner loop uses 38 flops */
 655         }
 656
 657         if(jidx<j_index_end)
 658         {
 659
 660             /* Get j neighbor index, and coordinate index */
 661             jnrlistA         = jjnr[jidx];
 662             jnrlistB         = jjnr[jidx+1];
 663             jnrlistC         = jjnr[jidx+2];
 664             jnrlistD         = jjnr[jidx+3];
 665             /* Sign of each element will be negative for non-real atoms.
 666              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 667              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 668              */
 669             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 670             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 671             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 672             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 673             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 674             j_coord_offsetA  = DIM*jnrA;
 675             j_coord_offsetB  = DIM*jnrB;
 676             j_coord_offsetC  = DIM*jnrC;
 677             j_coord_offsetD  = DIM*jnrD;
 678
 679             /* load j atom coordinates */
 680             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 681                                               x+j_coord_offsetC,x+j_coord_offsetD,
 682                                               &jx0,&jy0,&jz0);
 683
 684             /* Calculate displacement vector */
 685             dx00             = _mm_sub_ps(ix0,jx0);
 686             dy00             = _mm_sub_ps(iy0,jy0);
 687             dz00             = _mm_sub_ps(iz0,jz0);
 688
 689             /* Calculate squared distance and things based on it */
 690             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 691
 692             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 693
 694             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 695
 696             /* Load parameters for j particles */
 697             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 698                                                               charge+jnrC+0,charge+jnrD+0);
 699             vdwjidx0A        = 2*vdwtype[jnrA+0];
 700             vdwjidx0B        = 2*vdwtype[jnrB+0];
 701             vdwjidx0C        = 2*vdwtype[jnrC+0];
 702             vdwjidx0D        = 2*vdwtype[jnrD+0];
 703
 704             /**************************
 705              * CALCULATE INTERACTIONS *
 706              **************************/
 707
 708             if (gmx_mm_any_lt(rsq00,rcutoff2))
 709             {
 710
 711             r00              = _mm_mul_ps(rsq00,rinv00);
 712             r00              = _mm_andnot_ps(dummy_mask,r00);
 713
 714             /* Compute parameters for interactions between i and j atoms */
 715             qq00             = _mm_mul_ps(iq0,jq0);
 716             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 717                                          vdwparam+vdwioffset0+vdwjidx0B,
 718                                          vdwparam+vdwioffset0+vdwjidx0C,
 719                                          vdwparam+vdwioffset0+vdwjidx0D,
 720                                          &c6_00,&c12_00);
 721
 722             /* EWALD ELECTROSTATICS */
 723
 724             /* Analytical PME correction */
 725             zeta2            = _mm_mul_ps(beta2,rsq00);
 726             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 727             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 728             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 729             felec            = _mm_mul_ps(qq00,felec);
 730
 731             /* LENNARD-JONES DISPERSION/REPULSION */
 732
 733             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 734             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 735
 736             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 737
 738             fscal            = _mm_add_ps(felec,fvdw);
 739
 740             fscal            = _mm_and_ps(fscal,cutoff_mask);
 741
 742             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 743
 744              /* Update vectorial force */
 745             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 746             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 747             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 748
 749             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 750             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 751             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 752             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 753             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 754                                                    _mm_mul_ps(dx00,fscal),
 755                                                    _mm_mul_ps(dy00,fscal),
 756                                                    _mm_mul_ps(dz00,fscal));
 757
 758             }
 759
 760             /* Inner loop uses 39 flops */
 761         }
 762
 763         /* End of innermost loop */
 764
 765         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 766                                               f+i_coord_offset,fshift+i_shift_offset);
 767
 768         /* Increment number of inner iterations */
 769         inneriter                  += j_index_end - j_index_start;
 770
 771         /* Outer loop uses 7 flops */
 772     }
 773
 774     /* Increment number of outer iterations */
 775     outeriter        += nri;
 776
 777     /* Update outer/inner flops */
 778
 779     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*39);
 780 }