src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecEw_VdwLJ_GeomP1P1_avx_128_fma_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_single kernel generator.
  37  */
  38 #include "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_single.h"
  50 #include "kernelutil_x86_avx_128_fma_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJ_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_ElecEw_VdwLJ_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     /* Avoid stupid compiler warnings */
 132     jnrA = jnrB = jnrC = jnrD = 0;
 133     j_coord_offsetA = 0;
 134     j_coord_offsetB = 0;
 135     j_coord_offsetC = 0;
 136     j_coord_offsetD = 0;
 137
 138     outeriter        = 0;
 139     inneriter        = 0;
 140
 141     for(iidx=0;iidx<4*DIM;iidx++)
 142     {
 143         scratch[iidx] = 0.0;
 144     }
 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_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 162
 163         fix0             = _mm_setzero_ps();
 164         fiy0             = _mm_setzero_ps();
 165         fiz0             = _mm_setzero_ps();
 166
 167         /* Load parameters for i particles */
 168         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 169         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 170
 171         /* Reset potential sums */
 172         velecsum         = _mm_setzero_ps();
 173         vvdwsum          = _mm_setzero_ps();
 174
 175         /* Start inner kernel loop */
 176         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 177         {
 178
 179             /* Get j neighbor index, and coordinate index */
 180             jnrA             = jjnr[jidx];
 181             jnrB             = jjnr[jidx+1];
 182             jnrC             = jjnr[jidx+2];
 183             jnrD             = jjnr[jidx+3];
 184             j_coord_offsetA  = DIM*jnrA;
 185             j_coord_offsetB  = DIM*jnrB;
 186             j_coord_offsetC  = DIM*jnrC;
 187             j_coord_offsetD  = DIM*jnrD;
 188
 189             /* load j atom coordinates */
 190             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 191                                               x+j_coord_offsetC,x+j_coord_offsetD,
 192                                               &jx0,&jy0,&jz0);
 193
 194             /* Calculate displacement vector */
 195             dx00             = _mm_sub_ps(ix0,jx0);
 196             dy00             = _mm_sub_ps(iy0,jy0);
 197             dz00             = _mm_sub_ps(iz0,jz0);
 198
 199             /* Calculate squared distance and things based on it */
 200             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 201
 202             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 203
 204             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 205
 206             /* Load parameters for j particles */
 207             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 208                                                               charge+jnrC+0,charge+jnrD+0);
 209             vdwjidx0A        = 2*vdwtype[jnrA+0];
 210             vdwjidx0B        = 2*vdwtype[jnrB+0];
 211             vdwjidx0C        = 2*vdwtype[jnrC+0];
 212             vdwjidx0D        = 2*vdwtype[jnrD+0];
 213
 214             /**************************
 215              * CALCULATE INTERACTIONS *
 216              **************************/
 217
 218             r00              = _mm_mul_ps(rsq00,rinv00);
 219
 220             /* Compute parameters for interactions between i and j atoms */
 221             qq00             = _mm_mul_ps(iq0,jq0);
 222             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 223                                          vdwparam+vdwioffset0+vdwjidx0B,
 224                                          vdwparam+vdwioffset0+vdwjidx0C,
 225                                          vdwparam+vdwioffset0+vdwjidx0D,
 226                                          &c6_00,&c12_00);
 227
 228             /* EWALD ELECTROSTATICS */
 229
 230             /* Analytical PME correction */
 231             zeta2            = _mm_mul_ps(beta2,rsq00);
 232             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 233             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 234             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 235             felec            = _mm_mul_ps(qq00,felec);
 236             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 237             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 238             velec            = _mm_mul_ps(qq00,velec);
 239
 240             /* LENNARD-JONES DISPERSION/REPULSION */
 241
 242             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 243             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 244             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 245             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 246             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 247
 248             /* Update potential sum for this i atom from the interaction with this j atom. */
 249             velecsum         = _mm_add_ps(velecsum,velec);
 250             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 251
 252             fscal            = _mm_add_ps(felec,fvdw);
 253
 254              /* Update vectorial force */
 255             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 256             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 257             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 258
 259             fjptrA             = f+j_coord_offsetA;
 260             fjptrB             = f+j_coord_offsetB;
 261             fjptrC             = f+j_coord_offsetC;
 262             fjptrD             = f+j_coord_offsetD;
 263             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 264                                                    _mm_mul_ps(dx00,fscal),
 265                                                    _mm_mul_ps(dy00,fscal),
 266                                                    _mm_mul_ps(dz00,fscal));
 267
 268             /* Inner loop uses 41 flops */
 269         }
 270
 271         if(jidx<j_index_end)
 272         {
 273
 274             /* Get j neighbor index, and coordinate index */
 275             jnrlistA         = jjnr[jidx];
 276             jnrlistB         = jjnr[jidx+1];
 277             jnrlistC         = jjnr[jidx+2];
 278             jnrlistD         = jjnr[jidx+3];
 279             /* Sign of each element will be negative for non-real atoms.
 280              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 281              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 282              */
 283             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 284             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 285             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 286             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 287             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 288             j_coord_offsetA  = DIM*jnrA;
 289             j_coord_offsetB  = DIM*jnrB;
 290             j_coord_offsetC  = DIM*jnrC;
 291             j_coord_offsetD  = DIM*jnrD;
 292
 293             /* load j atom coordinates */
 294             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 295                                               x+j_coord_offsetC,x+j_coord_offsetD,
 296                                               &jx0,&jy0,&jz0);
 297
 298             /* Calculate displacement vector */
 299             dx00             = _mm_sub_ps(ix0,jx0);
 300             dy00             = _mm_sub_ps(iy0,jy0);
 301             dz00             = _mm_sub_ps(iz0,jz0);
 302
 303             /* Calculate squared distance and things based on it */
 304             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 305
 306             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 307
 308             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 309
 310             /* Load parameters for j particles */
 311             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 312                                                               charge+jnrC+0,charge+jnrD+0);
 313             vdwjidx0A        = 2*vdwtype[jnrA+0];
 314             vdwjidx0B        = 2*vdwtype[jnrB+0];
 315             vdwjidx0C        = 2*vdwtype[jnrC+0];
 316             vdwjidx0D        = 2*vdwtype[jnrD+0];
 317
 318             /**************************
 319              * CALCULATE INTERACTIONS *
 320              **************************/
 321
 322             r00              = _mm_mul_ps(rsq00,rinv00);
 323             r00              = _mm_andnot_ps(dummy_mask,r00);
 324
 325             /* Compute parameters for interactions between i and j atoms */
 326             qq00             = _mm_mul_ps(iq0,jq0);
 327             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 328                                          vdwparam+vdwioffset0+vdwjidx0B,
 329                                          vdwparam+vdwioffset0+vdwjidx0C,
 330                                          vdwparam+vdwioffset0+vdwjidx0D,
 331                                          &c6_00,&c12_00);
 332
 333             /* EWALD ELECTROSTATICS */
 334
 335             /* Analytical PME correction */
 336             zeta2            = _mm_mul_ps(beta2,rsq00);
 337             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 338             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 339             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 340             felec            = _mm_mul_ps(qq00,felec);
 341             pmecorrV         = gmx_mm_pmecorrV_ps(zeta2);
 342             velec            = _mm_nmacc_ps(pmecorrV,beta,rinv00);
 343             velec            = _mm_mul_ps(qq00,velec);
 344
 345             /* LENNARD-JONES DISPERSION/REPULSION */
 346
 347             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 348             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 349             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 350             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 351             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 352
 353             /* Update potential sum for this i atom from the interaction with this j atom. */
 354             velec            = _mm_andnot_ps(dummy_mask,velec);
 355             velecsum         = _mm_add_ps(velecsum,velec);
 356             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 357             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 358
 359             fscal            = _mm_add_ps(felec,fvdw);
 360
 361             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 362
 363              /* Update vectorial force */
 364             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 365             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 366             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 367
 368             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 369             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 370             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 371             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 372             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 373                                                    _mm_mul_ps(dx00,fscal),
 374                                                    _mm_mul_ps(dy00,fscal),
 375                                                    _mm_mul_ps(dz00,fscal));
 376
 377             /* Inner loop uses 42 flops */
 378         }
 379
 380         /* End of innermost loop */
 381
 382         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 383                                               f+i_coord_offset,fshift+i_shift_offset);
 384
 385         ggid                        = gid[iidx];
 386         /* Update potential energies */
 387         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 388         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 389
 390         /* Increment number of inner iterations */
 391         inneriter                  += j_index_end - j_index_start;
 392
 393         /* Outer loop uses 9 flops */
 394     }
 395
 396     /* Increment number of outer iterations */
 397     outeriter        += nri;
 398
 399     /* Update outer/inner flops */
 400
 401     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*42);
 402 }
 403 /*
 404  * Gromacs nonbonded kernel:   nb_kernel_ElecEw_VdwLJ_GeomP1P1_F_avx_128_fma_single
 405  * Electrostatics interaction: Ewald
 406  * VdW interaction:            LennardJones
 407  * Geometry:                   Particle-Particle
 408  * Calculate force/pot:        Force
 409  */
 410 void
 411 nb_kernel_ElecEw_VdwLJ_GeomP1P1_F_avx_128_fma_single
 412                     (t_nblist                    * gmx_restrict       nlist,
 413                      rvec                        * gmx_restrict          xx,
 414                      rvec                        * gmx_restrict          ff,
 415                      t_forcerec                  * gmx_restrict          fr,
 416                      t_mdatoms                   * gmx_restrict     mdatoms,
 417                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 418                      t_nrnb                      * gmx_restrict        nrnb)
 419 {
 420     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 421      * just 0 for non-waters.
 422      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 423      * jnr indices corresponding to data put in the four positions in the SIMD register.
 424      */
 425     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 426     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 427     int              jnrA,jnrB,jnrC,jnrD;
 428     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 429     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 430     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 431     real             rcutoff_scalar;
 432     real             *shiftvec,*fshift,*x,*f;
 433     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 434     real             scratch[4*DIM];
 435     __m128           fscal,rcutoff,rcutoff2,jidxall;
 436     int              vdwioffset0;
 437     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 438     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 439     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 440     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 441     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 442     real             *charge;
 443     int              nvdwtype;
 444     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 445     int              *vdwtype;
 446     real             *vdwparam;
 447     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 448     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 449     __m128i          ewitab;
 450     __m128           ewtabscale,eweps,twoeweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 451     __m128           beta,beta2,beta3,zeta2,pmecorrF,pmecorrV,rinv3;
 452     real             *ewtab;
 453     __m128           dummy_mask,cutoff_mask;
 454     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 455     __m128           one     = _mm_set1_ps(1.0);
 456     __m128           two     = _mm_set1_ps(2.0);
 457     x                = xx[0];
 458     f                = ff[0];
 459
 460     nri              = nlist->nri;
 461     iinr             = nlist->iinr;
 462     jindex           = nlist->jindex;
 463     jjnr             = nlist->jjnr;
 464     shiftidx         = nlist->shift;
 465     gid              = nlist->gid;
 466     shiftvec         = fr->shift_vec[0];
 467     fshift           = fr->fshift[0];
 468     facel            = _mm_set1_ps(fr->epsfac);
 469     charge           = mdatoms->chargeA;
 470     nvdwtype         = fr->ntype;
 471     vdwparam         = fr->nbfp;
 472     vdwtype          = mdatoms->typeA;
 473
 474     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 475     beta             = _mm_set1_ps(fr->ic->ewaldcoeff_q);
 476     beta2            = _mm_mul_ps(beta,beta);
 477     beta3            = _mm_mul_ps(beta,beta2);
 478     ewtab            = fr->ic->tabq_coul_F;
 479     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 480     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 481
 482     /* Avoid stupid compiler warnings */
 483     jnrA = jnrB = jnrC = jnrD = 0;
 484     j_coord_offsetA = 0;
 485     j_coord_offsetB = 0;
 486     j_coord_offsetC = 0;
 487     j_coord_offsetD = 0;
 488
 489     outeriter        = 0;
 490     inneriter        = 0;
 491
 492     for(iidx=0;iidx<4*DIM;iidx++)
 493     {
 494         scratch[iidx] = 0.0;
 495     }
 496
 497     /* Start outer loop over neighborlists */
 498     for(iidx=0; iidx<nri; iidx++)
 499     {
 500         /* Load shift vector for this list */
 501         i_shift_offset   = DIM*shiftidx[iidx];
 502
 503         /* Load limits for loop over neighbors */
 504         j_index_start    = jindex[iidx];
 505         j_index_end      = jindex[iidx+1];
 506
 507         /* Get outer coordinate index */
 508         inr              = iinr[iidx];
 509         i_coord_offset   = DIM*inr;
 510
 511         /* Load i particle coords and add shift vector */
 512         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 513
 514         fix0             = _mm_setzero_ps();
 515         fiy0             = _mm_setzero_ps();
 516         fiz0             = _mm_setzero_ps();
 517
 518         /* Load parameters for i particles */
 519         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 520         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 521
 522         /* Start inner kernel loop */
 523         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 524         {
 525
 526             /* Get j neighbor index, and coordinate index */
 527             jnrA             = jjnr[jidx];
 528             jnrB             = jjnr[jidx+1];
 529             jnrC             = jjnr[jidx+2];
 530             jnrD             = jjnr[jidx+3];
 531             j_coord_offsetA  = DIM*jnrA;
 532             j_coord_offsetB  = DIM*jnrB;
 533             j_coord_offsetC  = DIM*jnrC;
 534             j_coord_offsetD  = DIM*jnrD;
 535
 536             /* load j atom coordinates */
 537             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 538                                               x+j_coord_offsetC,x+j_coord_offsetD,
 539                                               &jx0,&jy0,&jz0);
 540
 541             /* Calculate displacement vector */
 542             dx00             = _mm_sub_ps(ix0,jx0);
 543             dy00             = _mm_sub_ps(iy0,jy0);
 544             dz00             = _mm_sub_ps(iz0,jz0);
 545
 546             /* Calculate squared distance and things based on it */
 547             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 548
 549             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 550
 551             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 552
 553             /* Load parameters for j particles */
 554             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 555                                                               charge+jnrC+0,charge+jnrD+0);
 556             vdwjidx0A        = 2*vdwtype[jnrA+0];
 557             vdwjidx0B        = 2*vdwtype[jnrB+0];
 558             vdwjidx0C        = 2*vdwtype[jnrC+0];
 559             vdwjidx0D        = 2*vdwtype[jnrD+0];
 560
 561             /**************************
 562              * CALCULATE INTERACTIONS *
 563              **************************/
 564
 565             r00              = _mm_mul_ps(rsq00,rinv00);
 566
 567             /* Compute parameters for interactions between i and j atoms */
 568             qq00             = _mm_mul_ps(iq0,jq0);
 569             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 570                                          vdwparam+vdwioffset0+vdwjidx0B,
 571                                          vdwparam+vdwioffset0+vdwjidx0C,
 572                                          vdwparam+vdwioffset0+vdwjidx0D,
 573                                          &c6_00,&c12_00);
 574
 575             /* EWALD ELECTROSTATICS */
 576
 577             /* Analytical PME correction */
 578             zeta2            = _mm_mul_ps(beta2,rsq00);
 579             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 580             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 581             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 582             felec            = _mm_mul_ps(qq00,felec);
 583
 584             /* LENNARD-JONES DISPERSION/REPULSION */
 585
 586             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 587             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 588
 589             fscal            = _mm_add_ps(felec,fvdw);
 590
 591              /* Update vectorial force */
 592             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 593             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 594             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 595
 596             fjptrA             = f+j_coord_offsetA;
 597             fjptrB             = f+j_coord_offsetB;
 598             fjptrC             = f+j_coord_offsetC;
 599             fjptrD             = f+j_coord_offsetD;
 600             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 601                                                    _mm_mul_ps(dx00,fscal),
 602                                                    _mm_mul_ps(dy00,fscal),
 603                                                    _mm_mul_ps(dz00,fscal));
 604
 605             /* Inner loop uses 35 flops */
 606         }
 607
 608         if(jidx<j_index_end)
 609         {
 610
 611             /* Get j neighbor index, and coordinate index */
 612             jnrlistA         = jjnr[jidx];
 613             jnrlistB         = jjnr[jidx+1];
 614             jnrlistC         = jjnr[jidx+2];
 615             jnrlistD         = jjnr[jidx+3];
 616             /* Sign of each element will be negative for non-real atoms.
 617              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 618              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 619              */
 620             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 621             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 622             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 623             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 624             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 625             j_coord_offsetA  = DIM*jnrA;
 626             j_coord_offsetB  = DIM*jnrB;
 627             j_coord_offsetC  = DIM*jnrC;
 628             j_coord_offsetD  = DIM*jnrD;
 629
 630             /* load j atom coordinates */
 631             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 632                                               x+j_coord_offsetC,x+j_coord_offsetD,
 633                                               &jx0,&jy0,&jz0);
 634
 635             /* Calculate displacement vector */
 636             dx00             = _mm_sub_ps(ix0,jx0);
 637             dy00             = _mm_sub_ps(iy0,jy0);
 638             dz00             = _mm_sub_ps(iz0,jz0);
 639
 640             /* Calculate squared distance and things based on it */
 641             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 642
 643             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 644
 645             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 646
 647             /* Load parameters for j particles */
 648             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 649                                                               charge+jnrC+0,charge+jnrD+0);
 650             vdwjidx0A        = 2*vdwtype[jnrA+0];
 651             vdwjidx0B        = 2*vdwtype[jnrB+0];
 652             vdwjidx0C        = 2*vdwtype[jnrC+0];
 653             vdwjidx0D        = 2*vdwtype[jnrD+0];
 654
 655             /**************************
 656              * CALCULATE INTERACTIONS *
 657              **************************/
 658
 659             r00              = _mm_mul_ps(rsq00,rinv00);
 660             r00              = _mm_andnot_ps(dummy_mask,r00);
 661
 662             /* Compute parameters for interactions between i and j atoms */
 663             qq00             = _mm_mul_ps(iq0,jq0);
 664             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 665                                          vdwparam+vdwioffset0+vdwjidx0B,
 666                                          vdwparam+vdwioffset0+vdwjidx0C,
 667                                          vdwparam+vdwioffset0+vdwjidx0D,
 668                                          &c6_00,&c12_00);
 669
 670             /* EWALD ELECTROSTATICS */
 671
 672             /* Analytical PME correction */
 673             zeta2            = _mm_mul_ps(beta2,rsq00);
 674             rinv3            = _mm_mul_ps(rinvsq00,rinv00);
 675             pmecorrF         = gmx_mm_pmecorrF_ps(zeta2);
 676             felec            = _mm_macc_ps(pmecorrF,beta3,rinv3);
 677             felec            = _mm_mul_ps(qq00,felec);
 678
 679             /* LENNARD-JONES DISPERSION/REPULSION */
 680
 681             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 682             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 683
 684             fscal            = _mm_add_ps(felec,fvdw);
 685
 686             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 687
 688              /* Update vectorial force */
 689             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 690             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 691             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 692
 693             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 694             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 695             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 696             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 697             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 698                                                    _mm_mul_ps(dx00,fscal),
 699                                                    _mm_mul_ps(dy00,fscal),
 700                                                    _mm_mul_ps(dz00,fscal));
 701
 702             /* Inner loop uses 36 flops */
 703         }
 704
 705         /* End of innermost loop */
 706
 707         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 708                                               f+i_coord_offset,fshift+i_shift_offset);
 709
 710         /* Increment number of inner iterations */
 711         inneriter                  += j_index_end - j_index_start;
 712
 713         /* Outer loop uses 7 flops */
 714     }
 715
 716     /* Increment number of outer iterations */
 717     outeriter        += nri;
 718
 719     /* Update outer/inner flops */
 720
 721     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*36);
 722 }