src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecCSTab_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 #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_ElecCSTab_VdwLJ_GeomP1P1_VF_avx_128_fma_single
  54  * Electrostatics interaction: CubicSplineTable
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecCSTab_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          vfitab;
  99     __m128i          ifour       = _mm_set1_epi32(4);
 100     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 101     real             *vftab;
 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     vftab            = kernel_data->table_elec->data;
 124     vftabscale       = _mm_set1_ps(kernel_data->table_elec->scale);
 125
 126     /* Avoid stupid compiler warnings */
 127     jnrA = jnrB = jnrC = jnrD = 0;
 128     j_coord_offsetA = 0;
 129     j_coord_offsetB = 0;
 130     j_coord_offsetC = 0;
 131     j_coord_offsetD = 0;
 132
 133     outeriter        = 0;
 134     inneriter        = 0;
 135
 136     for(iidx=0;iidx<4*DIM;iidx++)
 137     {
 138         scratch[iidx] = 0.0;
 139     }
 140
 141     /* Start outer loop over neighborlists */
 142     for(iidx=0; iidx<nri; iidx++)
 143     {
 144         /* Load shift vector for this list */
 145         i_shift_offset   = DIM*shiftidx[iidx];
 146
 147         /* Load limits for loop over neighbors */
 148         j_index_start    = jindex[iidx];
 149         j_index_end      = jindex[iidx+1];
 150
 151         /* Get outer coordinate index */
 152         inr              = iinr[iidx];
 153         i_coord_offset   = DIM*inr;
 154
 155         /* Load i particle coords and add shift vector */
 156         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 157
 158         fix0             = _mm_setzero_ps();
 159         fiy0             = _mm_setzero_ps();
 160         fiz0             = _mm_setzero_ps();
 161
 162         /* Load parameters for i particles */
 163         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 164         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 165
 166         /* Reset potential sums */
 167         velecsum         = _mm_setzero_ps();
 168         vvdwsum          = _mm_setzero_ps();
 169
 170         /* Start inner kernel loop */
 171         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 172         {
 173
 174             /* Get j neighbor index, and coordinate index */
 175             jnrA             = jjnr[jidx];
 176             jnrB             = jjnr[jidx+1];
 177             jnrC             = jjnr[jidx+2];
 178             jnrD             = jjnr[jidx+3];
 179             j_coord_offsetA  = DIM*jnrA;
 180             j_coord_offsetB  = DIM*jnrB;
 181             j_coord_offsetC  = DIM*jnrC;
 182             j_coord_offsetD  = DIM*jnrD;
 183
 184             /* load j atom coordinates */
 185             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 186                                               x+j_coord_offsetC,x+j_coord_offsetD,
 187                                               &jx0,&jy0,&jz0);
 188
 189             /* Calculate displacement vector */
 190             dx00             = _mm_sub_ps(ix0,jx0);
 191             dy00             = _mm_sub_ps(iy0,jy0);
 192             dz00             = _mm_sub_ps(iz0,jz0);
 193
 194             /* Calculate squared distance and things based on it */
 195             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 196
 197             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 198
 199             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 200
 201             /* Load parameters for j particles */
 202             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 203                                                               charge+jnrC+0,charge+jnrD+0);
 204             vdwjidx0A        = 2*vdwtype[jnrA+0];
 205             vdwjidx0B        = 2*vdwtype[jnrB+0];
 206             vdwjidx0C        = 2*vdwtype[jnrC+0];
 207             vdwjidx0D        = 2*vdwtype[jnrD+0];
 208
 209             /**************************
 210              * CALCULATE INTERACTIONS *
 211              **************************/
 212
 213             r00              = _mm_mul_ps(rsq00,rinv00);
 214
 215             /* Compute parameters for interactions between i and j atoms */
 216             qq00             = _mm_mul_ps(iq0,jq0);
 217             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 218                                          vdwparam+vdwioffset0+vdwjidx0B,
 219                                          vdwparam+vdwioffset0+vdwjidx0C,
 220                                          vdwparam+vdwioffset0+vdwjidx0D,
 221                                          &c6_00,&c12_00);
 222
 223             /* Calculate table index by multiplying r with table scale and truncate to integer */
 224             rt               = _mm_mul_ps(r00,vftabscale);
 225             vfitab           = _mm_cvttps_epi32(rt);
 226 #ifdef __XOP__
 227             vfeps            = _mm_frcz_ps(rt);
 228 #else
 229             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 230 #endif
 231             twovfeps         = _mm_add_ps(vfeps,vfeps);
 232             vfitab           = _mm_slli_epi32(vfitab,2);
 233
 234             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 235             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 236             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 237             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 238             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 239             _MM_TRANSPOSE4_PS(Y,F,G,H);
 240             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 241             VV               = _mm_macc_ps(vfeps,Fp,Y);
 242             velec            = _mm_mul_ps(qq00,VV);
 243             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 244             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 245
 246             /* LENNARD-JONES DISPERSION/REPULSION */
 247
 248             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 249             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 250             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 251             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 252             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 253
 254             /* Update potential sum for this i atom from the interaction with this j atom. */
 255             velecsum         = _mm_add_ps(velecsum,velec);
 256             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 257
 258             fscal            = _mm_add_ps(felec,fvdw);
 259
 260              /* Update vectorial force */
 261             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 262             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 263             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 264
 265             fjptrA             = f+j_coord_offsetA;
 266             fjptrB             = f+j_coord_offsetB;
 267             fjptrC             = f+j_coord_offsetC;
 268             fjptrD             = f+j_coord_offsetD;
 269             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 270                                                    _mm_mul_ps(dx00,fscal),
 271                                                    _mm_mul_ps(dy00,fscal),
 272                                                    _mm_mul_ps(dz00,fscal));
 273
 274             /* Inner loop uses 59 flops */
 275         }
 276
 277         if(jidx<j_index_end)
 278         {
 279
 280             /* Get j neighbor index, and coordinate index */
 281             jnrlistA         = jjnr[jidx];
 282             jnrlistB         = jjnr[jidx+1];
 283             jnrlistC         = jjnr[jidx+2];
 284             jnrlistD         = jjnr[jidx+3];
 285             /* Sign of each element will be negative for non-real atoms.
 286              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 287              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 288              */
 289             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 290             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 291             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 292             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 293             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 294             j_coord_offsetA  = DIM*jnrA;
 295             j_coord_offsetB  = DIM*jnrB;
 296             j_coord_offsetC  = DIM*jnrC;
 297             j_coord_offsetD  = DIM*jnrD;
 298
 299             /* load j atom coordinates */
 300             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 301                                               x+j_coord_offsetC,x+j_coord_offsetD,
 302                                               &jx0,&jy0,&jz0);
 303
 304             /* Calculate displacement vector */
 305             dx00             = _mm_sub_ps(ix0,jx0);
 306             dy00             = _mm_sub_ps(iy0,jy0);
 307             dz00             = _mm_sub_ps(iz0,jz0);
 308
 309             /* Calculate squared distance and things based on it */
 310             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 311
 312             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 313
 314             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 315
 316             /* Load parameters for j particles */
 317             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 318                                                               charge+jnrC+0,charge+jnrD+0);
 319             vdwjidx0A        = 2*vdwtype[jnrA+0];
 320             vdwjidx0B        = 2*vdwtype[jnrB+0];
 321             vdwjidx0C        = 2*vdwtype[jnrC+0];
 322             vdwjidx0D        = 2*vdwtype[jnrD+0];
 323
 324             /**************************
 325              * CALCULATE INTERACTIONS *
 326              **************************/
 327
 328             r00              = _mm_mul_ps(rsq00,rinv00);
 329             r00              = _mm_andnot_ps(dummy_mask,r00);
 330
 331             /* Compute parameters for interactions between i and j atoms */
 332             qq00             = _mm_mul_ps(iq0,jq0);
 333             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 334                                          vdwparam+vdwioffset0+vdwjidx0B,
 335                                          vdwparam+vdwioffset0+vdwjidx0C,
 336                                          vdwparam+vdwioffset0+vdwjidx0D,
 337                                          &c6_00,&c12_00);
 338
 339             /* Calculate table index by multiplying r with table scale and truncate to integer */
 340             rt               = _mm_mul_ps(r00,vftabscale);
 341             vfitab           = _mm_cvttps_epi32(rt);
 342 #ifdef __XOP__
 343             vfeps            = _mm_frcz_ps(rt);
 344 #else
 345             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 346 #endif
 347             twovfeps         = _mm_add_ps(vfeps,vfeps);
 348             vfitab           = _mm_slli_epi32(vfitab,2);
 349
 350             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 351             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 352             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 353             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 354             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 355             _MM_TRANSPOSE4_PS(Y,F,G,H);
 356             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 357             VV               = _mm_macc_ps(vfeps,Fp,Y);
 358             velec            = _mm_mul_ps(qq00,VV);
 359             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 360             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 361
 362             /* LENNARD-JONES DISPERSION/REPULSION */
 363
 364             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 365             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 366             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 367             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 368             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 369
 370             /* Update potential sum for this i atom from the interaction with this j atom. */
 371             velec            = _mm_andnot_ps(dummy_mask,velec);
 372             velecsum         = _mm_add_ps(velecsum,velec);
 373             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 374             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 375
 376             fscal            = _mm_add_ps(felec,fvdw);
 377
 378             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 379
 380              /* Update vectorial force */
 381             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 382             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 383             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 384
 385             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 386             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 387             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 388             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 389             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 390                                                    _mm_mul_ps(dx00,fscal),
 391                                                    _mm_mul_ps(dy00,fscal),
 392                                                    _mm_mul_ps(dz00,fscal));
 393
 394             /* Inner loop uses 60 flops */
 395         }
 396
 397         /* End of innermost loop */
 398
 399         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 400                                               f+i_coord_offset,fshift+i_shift_offset);
 401
 402         ggid                        = gid[iidx];
 403         /* Update potential energies */
 404         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 405         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 406
 407         /* Increment number of inner iterations */
 408         inneriter                  += j_index_end - j_index_start;
 409
 410         /* Outer loop uses 9 flops */
 411     }
 412
 413     /* Increment number of outer iterations */
 414     outeriter        += nri;
 415
 416     /* Update outer/inner flops */
 417
 418     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*60);
 419 }
 420 /*
 421  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwLJ_GeomP1P1_F_avx_128_fma_single
 422  * Electrostatics interaction: CubicSplineTable
 423  * VdW interaction:            LennardJones
 424  * Geometry:                   Particle-Particle
 425  * Calculate force/pot:        Force
 426  */
 427 void
 428 nb_kernel_ElecCSTab_VdwLJ_GeomP1P1_F_avx_128_fma_single
 429                     (t_nblist                    * gmx_restrict       nlist,
 430                      rvec                        * gmx_restrict          xx,
 431                      rvec                        * gmx_restrict          ff,
 432                      t_forcerec                  * gmx_restrict          fr,
 433                      t_mdatoms                   * gmx_restrict     mdatoms,
 434                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 435                      t_nrnb                      * gmx_restrict        nrnb)
 436 {
 437     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 438      * just 0 for non-waters.
 439      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 440      * jnr indices corresponding to data put in the four positions in the SIMD register.
 441      */
 442     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 443     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 444     int              jnrA,jnrB,jnrC,jnrD;
 445     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 446     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 447     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 448     real             rcutoff_scalar;
 449     real             *shiftvec,*fshift,*x,*f;
 450     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 451     real             scratch[4*DIM];
 452     __m128           fscal,rcutoff,rcutoff2,jidxall;
 453     int              vdwioffset0;
 454     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 455     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 456     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 457     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 458     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 459     real             *charge;
 460     int              nvdwtype;
 461     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 462     int              *vdwtype;
 463     real             *vdwparam;
 464     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 465     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 466     __m128i          vfitab;
 467     __m128i          ifour       = _mm_set1_epi32(4);
 468     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 469     real             *vftab;
 470     __m128           dummy_mask,cutoff_mask;
 471     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 472     __m128           one     = _mm_set1_ps(1.0);
 473     __m128           two     = _mm_set1_ps(2.0);
 474     x                = xx[0];
 475     f                = ff[0];
 476
 477     nri              = nlist->nri;
 478     iinr             = nlist->iinr;
 479     jindex           = nlist->jindex;
 480     jjnr             = nlist->jjnr;
 481     shiftidx         = nlist->shift;
 482     gid              = nlist->gid;
 483     shiftvec         = fr->shift_vec[0];
 484     fshift           = fr->fshift[0];
 485     facel            = _mm_set1_ps(fr->epsfac);
 486     charge           = mdatoms->chargeA;
 487     nvdwtype         = fr->ntype;
 488     vdwparam         = fr->nbfp;
 489     vdwtype          = mdatoms->typeA;
 490
 491     vftab            = kernel_data->table_elec->data;
 492     vftabscale       = _mm_set1_ps(kernel_data->table_elec->scale);
 493
 494     /* Avoid stupid compiler warnings */
 495     jnrA = jnrB = jnrC = jnrD = 0;
 496     j_coord_offsetA = 0;
 497     j_coord_offsetB = 0;
 498     j_coord_offsetC = 0;
 499     j_coord_offsetD = 0;
 500
 501     outeriter        = 0;
 502     inneriter        = 0;
 503
 504     for(iidx=0;iidx<4*DIM;iidx++)
 505     {
 506         scratch[iidx] = 0.0;
 507     }
 508
 509     /* Start outer loop over neighborlists */
 510     for(iidx=0; iidx<nri; iidx++)
 511     {
 512         /* Load shift vector for this list */
 513         i_shift_offset   = DIM*shiftidx[iidx];
 514
 515         /* Load limits for loop over neighbors */
 516         j_index_start    = jindex[iidx];
 517         j_index_end      = jindex[iidx+1];
 518
 519         /* Get outer coordinate index */
 520         inr              = iinr[iidx];
 521         i_coord_offset   = DIM*inr;
 522
 523         /* Load i particle coords and add shift vector */
 524         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 525
 526         fix0             = _mm_setzero_ps();
 527         fiy0             = _mm_setzero_ps();
 528         fiz0             = _mm_setzero_ps();
 529
 530         /* Load parameters for i particles */
 531         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 532         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 533
 534         /* Start inner kernel loop */
 535         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 536         {
 537
 538             /* Get j neighbor index, and coordinate index */
 539             jnrA             = jjnr[jidx];
 540             jnrB             = jjnr[jidx+1];
 541             jnrC             = jjnr[jidx+2];
 542             jnrD             = jjnr[jidx+3];
 543             j_coord_offsetA  = DIM*jnrA;
 544             j_coord_offsetB  = DIM*jnrB;
 545             j_coord_offsetC  = DIM*jnrC;
 546             j_coord_offsetD  = DIM*jnrD;
 547
 548             /* load j atom coordinates */
 549             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 550                                               x+j_coord_offsetC,x+j_coord_offsetD,
 551                                               &jx0,&jy0,&jz0);
 552
 553             /* Calculate displacement vector */
 554             dx00             = _mm_sub_ps(ix0,jx0);
 555             dy00             = _mm_sub_ps(iy0,jy0);
 556             dz00             = _mm_sub_ps(iz0,jz0);
 557
 558             /* Calculate squared distance and things based on it */
 559             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 560
 561             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 562
 563             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 564
 565             /* Load parameters for j particles */
 566             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 567                                                               charge+jnrC+0,charge+jnrD+0);
 568             vdwjidx0A        = 2*vdwtype[jnrA+0];
 569             vdwjidx0B        = 2*vdwtype[jnrB+0];
 570             vdwjidx0C        = 2*vdwtype[jnrC+0];
 571             vdwjidx0D        = 2*vdwtype[jnrD+0];
 572
 573             /**************************
 574              * CALCULATE INTERACTIONS *
 575              **************************/
 576
 577             r00              = _mm_mul_ps(rsq00,rinv00);
 578
 579             /* Compute parameters for interactions between i and j atoms */
 580             qq00             = _mm_mul_ps(iq0,jq0);
 581             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 582                                          vdwparam+vdwioffset0+vdwjidx0B,
 583                                          vdwparam+vdwioffset0+vdwjidx0C,
 584                                          vdwparam+vdwioffset0+vdwjidx0D,
 585                                          &c6_00,&c12_00);
 586
 587             /* Calculate table index by multiplying r with table scale and truncate to integer */
 588             rt               = _mm_mul_ps(r00,vftabscale);
 589             vfitab           = _mm_cvttps_epi32(rt);
 590 #ifdef __XOP__
 591             vfeps            = _mm_frcz_ps(rt);
 592 #else
 593             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 594 #endif
 595             twovfeps         = _mm_add_ps(vfeps,vfeps);
 596             vfitab           = _mm_slli_epi32(vfitab,2);
 597
 598             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 599             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 600             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 601             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 602             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 603             _MM_TRANSPOSE4_PS(Y,F,G,H);
 604             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 605             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 606             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 607
 608             /* LENNARD-JONES DISPERSION/REPULSION */
 609
 610             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 611             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 612
 613             fscal            = _mm_add_ps(felec,fvdw);
 614
 615              /* Update vectorial force */
 616             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 617             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 618             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 619
 620             fjptrA             = f+j_coord_offsetA;
 621             fjptrB             = f+j_coord_offsetB;
 622             fjptrC             = f+j_coord_offsetC;
 623             fjptrD             = f+j_coord_offsetD;
 624             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 625                                                    _mm_mul_ps(dx00,fscal),
 626                                                    _mm_mul_ps(dy00,fscal),
 627                                                    _mm_mul_ps(dz00,fscal));
 628
 629             /* Inner loop uses 50 flops */
 630         }
 631
 632         if(jidx<j_index_end)
 633         {
 634
 635             /* Get j neighbor index, and coordinate index */
 636             jnrlistA         = jjnr[jidx];
 637             jnrlistB         = jjnr[jidx+1];
 638             jnrlistC         = jjnr[jidx+2];
 639             jnrlistD         = jjnr[jidx+3];
 640             /* Sign of each element will be negative for non-real atoms.
 641              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 642              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 643              */
 644             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 645             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 646             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 647             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 648             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 649             j_coord_offsetA  = DIM*jnrA;
 650             j_coord_offsetB  = DIM*jnrB;
 651             j_coord_offsetC  = DIM*jnrC;
 652             j_coord_offsetD  = DIM*jnrD;
 653
 654             /* load j atom coordinates */
 655             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 656                                               x+j_coord_offsetC,x+j_coord_offsetD,
 657                                               &jx0,&jy0,&jz0);
 658
 659             /* Calculate displacement vector */
 660             dx00             = _mm_sub_ps(ix0,jx0);
 661             dy00             = _mm_sub_ps(iy0,jy0);
 662             dz00             = _mm_sub_ps(iz0,jz0);
 663
 664             /* Calculate squared distance and things based on it */
 665             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 666
 667             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 668
 669             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 670
 671             /* Load parameters for j particles */
 672             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 673                                                               charge+jnrC+0,charge+jnrD+0);
 674             vdwjidx0A        = 2*vdwtype[jnrA+0];
 675             vdwjidx0B        = 2*vdwtype[jnrB+0];
 676             vdwjidx0C        = 2*vdwtype[jnrC+0];
 677             vdwjidx0D        = 2*vdwtype[jnrD+0];
 678
 679             /**************************
 680              * CALCULATE INTERACTIONS *
 681              **************************/
 682
 683             r00              = _mm_mul_ps(rsq00,rinv00);
 684             r00              = _mm_andnot_ps(dummy_mask,r00);
 685
 686             /* Compute parameters for interactions between i and j atoms */
 687             qq00             = _mm_mul_ps(iq0,jq0);
 688             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 689                                          vdwparam+vdwioffset0+vdwjidx0B,
 690                                          vdwparam+vdwioffset0+vdwjidx0C,
 691                                          vdwparam+vdwioffset0+vdwjidx0D,
 692                                          &c6_00,&c12_00);
 693
 694             /* Calculate table index by multiplying r with table scale and truncate to integer */
 695             rt               = _mm_mul_ps(r00,vftabscale);
 696             vfitab           = _mm_cvttps_epi32(rt);
 697 #ifdef __XOP__
 698             vfeps            = _mm_frcz_ps(rt);
 699 #else
 700             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 701 #endif
 702             twovfeps         = _mm_add_ps(vfeps,vfeps);
 703             vfitab           = _mm_slli_epi32(vfitab,2);
 704
 705             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 706             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 707             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 708             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 709             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 710             _MM_TRANSPOSE4_PS(Y,F,G,H);
 711             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 712             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 713             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 714
 715             /* LENNARD-JONES DISPERSION/REPULSION */
 716
 717             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 718             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 719
 720             fscal            = _mm_add_ps(felec,fvdw);
 721
 722             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 723
 724              /* Update vectorial force */
 725             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 726             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 727             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 728
 729             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 730             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 731             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 732             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 733             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 734                                                    _mm_mul_ps(dx00,fscal),
 735                                                    _mm_mul_ps(dy00,fscal),
 736                                                    _mm_mul_ps(dz00,fscal));
 737
 738             /* Inner loop uses 51 flops */
 739         }
 740
 741         /* End of innermost loop */
 742
 743         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 744                                               f+i_coord_offset,fshift+i_shift_offset);
 745
 746         /* Increment number of inner iterations */
 747         inneriter                  += j_index_end - j_index_start;
 748
 749         /* Outer loop uses 7 flops */
 750     }
 751
 752     /* Increment number of outer iterations */
 753     outeriter        += nri;
 754
 755     /* Update outer/inner flops */
 756
 757     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*51);
 758 }