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 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  48 #include "kernelutil_x86_avx_128_fma_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwLJ_GeomP1P1_VF_avx_128_fma_single
  52  * Electrostatics interaction: CubicSplineTable
  53  * VdW interaction:            LennardJones
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecCSTab_VdwLJ_GeomP1P1_VF_avx_128_fma_single
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m128           fscal,rcutoff,rcutoff2,jidxall;
  83     int              vdwioffset0;
  84     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  85     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  86     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  87     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  88     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  89     real             *charge;
  90     int              nvdwtype;
  91     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  92     int              *vdwtype;
  93     real             *vdwparam;
  94     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  95     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  96     __m128i          vfitab;
  97     __m128i          ifour       = _mm_set1_epi32(4);
  98     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
  99     real             *vftab;
 100     __m128           dummy_mask,cutoff_mask;
 101     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 102     __m128           one     = _mm_set1_ps(1.0);
 103     __m128           two     = _mm_set1_ps(2.0);
 104     x                = xx[0];
 105     f                = ff[0];
 106
 107     nri              = nlist->nri;
 108     iinr             = nlist->iinr;
 109     jindex           = nlist->jindex;
 110     jjnr             = nlist->jjnr;
 111     shiftidx         = nlist->shift;
 112     gid              = nlist->gid;
 113     shiftvec         = fr->shift_vec[0];
 114     fshift           = fr->fshift[0];
 115     facel            = _mm_set1_ps(fr->epsfac);
 116     charge           = mdatoms->chargeA;
 117     nvdwtype         = fr->ntype;
 118     vdwparam         = fr->nbfp;
 119     vdwtype          = mdatoms->typeA;
 120
 121     vftab            = kernel_data->table_elec->data;
 122     vftabscale       = _mm_set1_ps(kernel_data->table_elec->scale);
 123
 124     /* Avoid stupid compiler warnings */
 125     jnrA = jnrB = jnrC = jnrD = 0;
 126     j_coord_offsetA = 0;
 127     j_coord_offsetB = 0;
 128     j_coord_offsetC = 0;
 129     j_coord_offsetD = 0;
 130
 131     outeriter        = 0;
 132     inneriter        = 0;
 133
 134     for(iidx=0;iidx<4*DIM;iidx++)
 135     {
 136         scratch[iidx] = 0.0;
 137     }
 138
 139     /* Start outer loop over neighborlists */
 140     for(iidx=0; iidx<nri; iidx++)
 141     {
 142         /* Load shift vector for this list */
 143         i_shift_offset   = DIM*shiftidx[iidx];
 144
 145         /* Load limits for loop over neighbors */
 146         j_index_start    = jindex[iidx];
 147         j_index_end      = jindex[iidx+1];
 148
 149         /* Get outer coordinate index */
 150         inr              = iinr[iidx];
 151         i_coord_offset   = DIM*inr;
 152
 153         /* Load i particle coords and add shift vector */
 154         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 155
 156         fix0             = _mm_setzero_ps();
 157         fiy0             = _mm_setzero_ps();
 158         fiz0             = _mm_setzero_ps();
 159
 160         /* Load parameters for i particles */
 161         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 162         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 163
 164         /* Reset potential sums */
 165         velecsum         = _mm_setzero_ps();
 166         vvdwsum          = _mm_setzero_ps();
 167
 168         /* Start inner kernel loop */
 169         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 170         {
 171
 172             /* Get j neighbor index, and coordinate index */
 173             jnrA             = jjnr[jidx];
 174             jnrB             = jjnr[jidx+1];
 175             jnrC             = jjnr[jidx+2];
 176             jnrD             = jjnr[jidx+3];
 177             j_coord_offsetA  = DIM*jnrA;
 178             j_coord_offsetB  = DIM*jnrB;
 179             j_coord_offsetC  = DIM*jnrC;
 180             j_coord_offsetD  = DIM*jnrD;
 181
 182             /* load j atom coordinates */
 183             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 184                                               x+j_coord_offsetC,x+j_coord_offsetD,
 185                                               &jx0,&jy0,&jz0);
 186
 187             /* Calculate displacement vector */
 188             dx00             = _mm_sub_ps(ix0,jx0);
 189             dy00             = _mm_sub_ps(iy0,jy0);
 190             dz00             = _mm_sub_ps(iz0,jz0);
 191
 192             /* Calculate squared distance and things based on it */
 193             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 194
 195             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 196
 197             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 198
 199             /* Load parameters for j particles */
 200             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 201                                                               charge+jnrC+0,charge+jnrD+0);
 202             vdwjidx0A        = 2*vdwtype[jnrA+0];
 203             vdwjidx0B        = 2*vdwtype[jnrB+0];
 204             vdwjidx0C        = 2*vdwtype[jnrC+0];
 205             vdwjidx0D        = 2*vdwtype[jnrD+0];
 206
 207             /**************************
 208              * CALCULATE INTERACTIONS *
 209              **************************/
 210
 211             r00              = _mm_mul_ps(rsq00,rinv00);
 212
 213             /* Compute parameters for interactions between i and j atoms */
 214             qq00             = _mm_mul_ps(iq0,jq0);
 215             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 216                                          vdwparam+vdwioffset0+vdwjidx0B,
 217                                          vdwparam+vdwioffset0+vdwjidx0C,
 218                                          vdwparam+vdwioffset0+vdwjidx0D,
 219                                          &c6_00,&c12_00);
 220
 221             /* Calculate table index by multiplying r with table scale and truncate to integer */
 222             rt               = _mm_mul_ps(r00,vftabscale);
 223             vfitab           = _mm_cvttps_epi32(rt);
 224 #ifdef __XOP__
 225             vfeps            = _mm_frcz_ps(rt);
 226 #else
 227             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 228 #endif
 229             twovfeps         = _mm_add_ps(vfeps,vfeps);
 230             vfitab           = _mm_slli_epi32(vfitab,2);
 231
 232             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 233             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 234             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 235             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 236             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 237             _MM_TRANSPOSE4_PS(Y,F,G,H);
 238             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 239             VV               = _mm_macc_ps(vfeps,Fp,Y);
 240             velec            = _mm_mul_ps(qq00,VV);
 241             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 242             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 243
 244             /* LENNARD-JONES DISPERSION/REPULSION */
 245
 246             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 247             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 248             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 249             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 250             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 251
 252             /* Update potential sum for this i atom from the interaction with this j atom. */
 253             velecsum         = _mm_add_ps(velecsum,velec);
 254             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 255
 256             fscal            = _mm_add_ps(felec,fvdw);
 257
 258              /* Update vectorial force */
 259             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 260             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 261             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 262
 263             fjptrA             = f+j_coord_offsetA;
 264             fjptrB             = f+j_coord_offsetB;
 265             fjptrC             = f+j_coord_offsetC;
 266             fjptrD             = f+j_coord_offsetD;
 267             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 268                                                    _mm_mul_ps(dx00,fscal),
 269                                                    _mm_mul_ps(dy00,fscal),
 270                                                    _mm_mul_ps(dz00,fscal));
 271
 272             /* Inner loop uses 59 flops */
 273         }
 274
 275         if(jidx<j_index_end)
 276         {
 277
 278             /* Get j neighbor index, and coordinate index */
 279             jnrlistA         = jjnr[jidx];
 280             jnrlistB         = jjnr[jidx+1];
 281             jnrlistC         = jjnr[jidx+2];
 282             jnrlistD         = jjnr[jidx+3];
 283             /* Sign of each element will be negative for non-real atoms.
 284              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 285              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 286              */
 287             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 288             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 289             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 290             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 291             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 292             j_coord_offsetA  = DIM*jnrA;
 293             j_coord_offsetB  = DIM*jnrB;
 294             j_coord_offsetC  = DIM*jnrC;
 295             j_coord_offsetD  = DIM*jnrD;
 296
 297             /* load j atom coordinates */
 298             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 299                                               x+j_coord_offsetC,x+j_coord_offsetD,
 300                                               &jx0,&jy0,&jz0);
 301
 302             /* Calculate displacement vector */
 303             dx00             = _mm_sub_ps(ix0,jx0);
 304             dy00             = _mm_sub_ps(iy0,jy0);
 305             dz00             = _mm_sub_ps(iz0,jz0);
 306
 307             /* Calculate squared distance and things based on it */
 308             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 309
 310             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 311
 312             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 313
 314             /* Load parameters for j particles */
 315             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 316                                                               charge+jnrC+0,charge+jnrD+0);
 317             vdwjidx0A        = 2*vdwtype[jnrA+0];
 318             vdwjidx0B        = 2*vdwtype[jnrB+0];
 319             vdwjidx0C        = 2*vdwtype[jnrC+0];
 320             vdwjidx0D        = 2*vdwtype[jnrD+0];
 321
 322             /**************************
 323              * CALCULATE INTERACTIONS *
 324              **************************/
 325
 326             r00              = _mm_mul_ps(rsq00,rinv00);
 327             r00              = _mm_andnot_ps(dummy_mask,r00);
 328
 329             /* Compute parameters for interactions between i and j atoms */
 330             qq00             = _mm_mul_ps(iq0,jq0);
 331             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 332                                          vdwparam+vdwioffset0+vdwjidx0B,
 333                                          vdwparam+vdwioffset0+vdwjidx0C,
 334                                          vdwparam+vdwioffset0+vdwjidx0D,
 335                                          &c6_00,&c12_00);
 336
 337             /* Calculate table index by multiplying r with table scale and truncate to integer */
 338             rt               = _mm_mul_ps(r00,vftabscale);
 339             vfitab           = _mm_cvttps_epi32(rt);
 340 #ifdef __XOP__
 341             vfeps            = _mm_frcz_ps(rt);
 342 #else
 343             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 344 #endif
 345             twovfeps         = _mm_add_ps(vfeps,vfeps);
 346             vfitab           = _mm_slli_epi32(vfitab,2);
 347
 348             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 349             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 350             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 351             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 352             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 353             _MM_TRANSPOSE4_PS(Y,F,G,H);
 354             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 355             VV               = _mm_macc_ps(vfeps,Fp,Y);
 356             velec            = _mm_mul_ps(qq00,VV);
 357             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 358             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 359
 360             /* LENNARD-JONES DISPERSION/REPULSION */
 361
 362             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 363             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 364             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 365             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 366             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 367
 368             /* Update potential sum for this i atom from the interaction with this j atom. */
 369             velec            = _mm_andnot_ps(dummy_mask,velec);
 370             velecsum         = _mm_add_ps(velecsum,velec);
 371             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 372             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 373
 374             fscal            = _mm_add_ps(felec,fvdw);
 375
 376             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 377
 378              /* Update vectorial force */
 379             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 380             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 381             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 382
 383             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 384             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 385             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 386             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 387             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 388                                                    _mm_mul_ps(dx00,fscal),
 389                                                    _mm_mul_ps(dy00,fscal),
 390                                                    _mm_mul_ps(dz00,fscal));
 391
 392             /* Inner loop uses 60 flops */
 393         }
 394
 395         /* End of innermost loop */
 396
 397         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 398                                               f+i_coord_offset,fshift+i_shift_offset);
 399
 400         ggid                        = gid[iidx];
 401         /* Update potential energies */
 402         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 403         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 404
 405         /* Increment number of inner iterations */
 406         inneriter                  += j_index_end - j_index_start;
 407
 408         /* Outer loop uses 9 flops */
 409     }
 410
 411     /* Increment number of outer iterations */
 412     outeriter        += nri;
 413
 414     /* Update outer/inner flops */
 415
 416     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*60);
 417 }
 418 /*
 419  * Gromacs nonbonded kernel:   nb_kernel_ElecCSTab_VdwLJ_GeomP1P1_F_avx_128_fma_single
 420  * Electrostatics interaction: CubicSplineTable
 421  * VdW interaction:            LennardJones
 422  * Geometry:                   Particle-Particle
 423  * Calculate force/pot:        Force
 424  */
 425 void
 426 nb_kernel_ElecCSTab_VdwLJ_GeomP1P1_F_avx_128_fma_single
 427                     (t_nblist                    * gmx_restrict       nlist,
 428                      rvec                        * gmx_restrict          xx,
 429                      rvec                        * gmx_restrict          ff,
 430                      t_forcerec                  * gmx_restrict          fr,
 431                      t_mdatoms                   * gmx_restrict     mdatoms,
 432                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 433                      t_nrnb                      * gmx_restrict        nrnb)
 434 {
 435     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 436      * just 0 for non-waters.
 437      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 438      * jnr indices corresponding to data put in the four positions in the SIMD register.
 439      */
 440     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 441     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 442     int              jnrA,jnrB,jnrC,jnrD;
 443     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 444     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 445     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 446     real             rcutoff_scalar;
 447     real             *shiftvec,*fshift,*x,*f;
 448     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 449     real             scratch[4*DIM];
 450     __m128           fscal,rcutoff,rcutoff2,jidxall;
 451     int              vdwioffset0;
 452     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 453     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 454     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 455     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 456     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 457     real             *charge;
 458     int              nvdwtype;
 459     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 460     int              *vdwtype;
 461     real             *vdwparam;
 462     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 463     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 464     __m128i          vfitab;
 465     __m128i          ifour       = _mm_set1_epi32(4);
 466     __m128           rt,vfeps,twovfeps,vftabscale,Y,F,G,H,Fp,VV,FF;
 467     real             *vftab;
 468     __m128           dummy_mask,cutoff_mask;
 469     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 470     __m128           one     = _mm_set1_ps(1.0);
 471     __m128           two     = _mm_set1_ps(2.0);
 472     x                = xx[0];
 473     f                = ff[0];
 474
 475     nri              = nlist->nri;
 476     iinr             = nlist->iinr;
 477     jindex           = nlist->jindex;
 478     jjnr             = nlist->jjnr;
 479     shiftidx         = nlist->shift;
 480     gid              = nlist->gid;
 481     shiftvec         = fr->shift_vec[0];
 482     fshift           = fr->fshift[0];
 483     facel            = _mm_set1_ps(fr->epsfac);
 484     charge           = mdatoms->chargeA;
 485     nvdwtype         = fr->ntype;
 486     vdwparam         = fr->nbfp;
 487     vdwtype          = mdatoms->typeA;
 488
 489     vftab            = kernel_data->table_elec->data;
 490     vftabscale       = _mm_set1_ps(kernel_data->table_elec->scale);
 491
 492     /* Avoid stupid compiler warnings */
 493     jnrA = jnrB = jnrC = jnrD = 0;
 494     j_coord_offsetA = 0;
 495     j_coord_offsetB = 0;
 496     j_coord_offsetC = 0;
 497     j_coord_offsetD = 0;
 498
 499     outeriter        = 0;
 500     inneriter        = 0;
 501
 502     for(iidx=0;iidx<4*DIM;iidx++)
 503     {
 504         scratch[iidx] = 0.0;
 505     }
 506
 507     /* Start outer loop over neighborlists */
 508     for(iidx=0; iidx<nri; iidx++)
 509     {
 510         /* Load shift vector for this list */
 511         i_shift_offset   = DIM*shiftidx[iidx];
 512
 513         /* Load limits for loop over neighbors */
 514         j_index_start    = jindex[iidx];
 515         j_index_end      = jindex[iidx+1];
 516
 517         /* Get outer coordinate index */
 518         inr              = iinr[iidx];
 519         i_coord_offset   = DIM*inr;
 520
 521         /* Load i particle coords and add shift vector */
 522         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 523
 524         fix0             = _mm_setzero_ps();
 525         fiy0             = _mm_setzero_ps();
 526         fiz0             = _mm_setzero_ps();
 527
 528         /* Load parameters for i particles */
 529         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 530         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 531
 532         /* Start inner kernel loop */
 533         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 534         {
 535
 536             /* Get j neighbor index, and coordinate index */
 537             jnrA             = jjnr[jidx];
 538             jnrB             = jjnr[jidx+1];
 539             jnrC             = jjnr[jidx+2];
 540             jnrD             = jjnr[jidx+3];
 541             j_coord_offsetA  = DIM*jnrA;
 542             j_coord_offsetB  = DIM*jnrB;
 543             j_coord_offsetC  = DIM*jnrC;
 544             j_coord_offsetD  = DIM*jnrD;
 545
 546             /* load j atom coordinates */
 547             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 548                                               x+j_coord_offsetC,x+j_coord_offsetD,
 549                                               &jx0,&jy0,&jz0);
 550
 551             /* Calculate displacement vector */
 552             dx00             = _mm_sub_ps(ix0,jx0);
 553             dy00             = _mm_sub_ps(iy0,jy0);
 554             dz00             = _mm_sub_ps(iz0,jz0);
 555
 556             /* Calculate squared distance and things based on it */
 557             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 558
 559             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 560
 561             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 562
 563             /* Load parameters for j particles */
 564             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 565                                                               charge+jnrC+0,charge+jnrD+0);
 566             vdwjidx0A        = 2*vdwtype[jnrA+0];
 567             vdwjidx0B        = 2*vdwtype[jnrB+0];
 568             vdwjidx0C        = 2*vdwtype[jnrC+0];
 569             vdwjidx0D        = 2*vdwtype[jnrD+0];
 570
 571             /**************************
 572              * CALCULATE INTERACTIONS *
 573              **************************/
 574
 575             r00              = _mm_mul_ps(rsq00,rinv00);
 576
 577             /* Compute parameters for interactions between i and j atoms */
 578             qq00             = _mm_mul_ps(iq0,jq0);
 579             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 580                                          vdwparam+vdwioffset0+vdwjidx0B,
 581                                          vdwparam+vdwioffset0+vdwjidx0C,
 582                                          vdwparam+vdwioffset0+vdwjidx0D,
 583                                          &c6_00,&c12_00);
 584
 585             /* Calculate table index by multiplying r with table scale and truncate to integer */
 586             rt               = _mm_mul_ps(r00,vftabscale);
 587             vfitab           = _mm_cvttps_epi32(rt);
 588 #ifdef __XOP__
 589             vfeps            = _mm_frcz_ps(rt);
 590 #else
 591             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 592 #endif
 593             twovfeps         = _mm_add_ps(vfeps,vfeps);
 594             vfitab           = _mm_slli_epi32(vfitab,2);
 595
 596             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 597             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 598             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 599             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 600             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 601             _MM_TRANSPOSE4_PS(Y,F,G,H);
 602             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 603             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 604             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 605
 606             /* LENNARD-JONES DISPERSION/REPULSION */
 607
 608             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 609             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 610
 611             fscal            = _mm_add_ps(felec,fvdw);
 612
 613              /* Update vectorial force */
 614             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 615             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 616             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 617
 618             fjptrA             = f+j_coord_offsetA;
 619             fjptrB             = f+j_coord_offsetB;
 620             fjptrC             = f+j_coord_offsetC;
 621             fjptrD             = f+j_coord_offsetD;
 622             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 623                                                    _mm_mul_ps(dx00,fscal),
 624                                                    _mm_mul_ps(dy00,fscal),
 625                                                    _mm_mul_ps(dz00,fscal));
 626
 627             /* Inner loop uses 50 flops */
 628         }
 629
 630         if(jidx<j_index_end)
 631         {
 632
 633             /* Get j neighbor index, and coordinate index */
 634             jnrlistA         = jjnr[jidx];
 635             jnrlistB         = jjnr[jidx+1];
 636             jnrlistC         = jjnr[jidx+2];
 637             jnrlistD         = jjnr[jidx+3];
 638             /* Sign of each element will be negative for non-real atoms.
 639              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 640              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 641              */
 642             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 643             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 644             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 645             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 646             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 647             j_coord_offsetA  = DIM*jnrA;
 648             j_coord_offsetB  = DIM*jnrB;
 649             j_coord_offsetC  = DIM*jnrC;
 650             j_coord_offsetD  = DIM*jnrD;
 651
 652             /* load j atom coordinates */
 653             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 654                                               x+j_coord_offsetC,x+j_coord_offsetD,
 655                                               &jx0,&jy0,&jz0);
 656
 657             /* Calculate displacement vector */
 658             dx00             = _mm_sub_ps(ix0,jx0);
 659             dy00             = _mm_sub_ps(iy0,jy0);
 660             dz00             = _mm_sub_ps(iz0,jz0);
 661
 662             /* Calculate squared distance and things based on it */
 663             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 664
 665             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 666
 667             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 668
 669             /* Load parameters for j particles */
 670             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 671                                                               charge+jnrC+0,charge+jnrD+0);
 672             vdwjidx0A        = 2*vdwtype[jnrA+0];
 673             vdwjidx0B        = 2*vdwtype[jnrB+0];
 674             vdwjidx0C        = 2*vdwtype[jnrC+0];
 675             vdwjidx0D        = 2*vdwtype[jnrD+0];
 676
 677             /**************************
 678              * CALCULATE INTERACTIONS *
 679              **************************/
 680
 681             r00              = _mm_mul_ps(rsq00,rinv00);
 682             r00              = _mm_andnot_ps(dummy_mask,r00);
 683
 684             /* Compute parameters for interactions between i and j atoms */
 685             qq00             = _mm_mul_ps(iq0,jq0);
 686             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 687                                          vdwparam+vdwioffset0+vdwjidx0B,
 688                                          vdwparam+vdwioffset0+vdwjidx0C,
 689                                          vdwparam+vdwioffset0+vdwjidx0D,
 690                                          &c6_00,&c12_00);
 691
 692             /* Calculate table index by multiplying r with table scale and truncate to integer */
 693             rt               = _mm_mul_ps(r00,vftabscale);
 694             vfitab           = _mm_cvttps_epi32(rt);
 695 #ifdef __XOP__
 696             vfeps            = _mm_frcz_ps(rt);
 697 #else
 698             vfeps            = _mm_sub_ps(rt,_mm_round_ps(rt, _MM_FROUND_FLOOR));
 699 #endif
 700             twovfeps         = _mm_add_ps(vfeps,vfeps);
 701             vfitab           = _mm_slli_epi32(vfitab,2);
 702
 703             /* CUBIC SPLINE TABLE ELECTROSTATICS */
 704             Y                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,0) );
 705             F                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,1) );
 706             G                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,2) );
 707             H                = _mm_load_ps( vftab + _mm_extract_epi32(vfitab,3) );
 708             _MM_TRANSPOSE4_PS(Y,F,G,H);
 709             Fp               = _mm_macc_ps(vfeps,_mm_macc_ps(H,vfeps,G),F);
 710             FF               = _mm_macc_ps(vfeps,_mm_macc_ps(twovfeps,H,G),Fp);
 711             felec            = _mm_xor_ps(signbit,_mm_mul_ps(_mm_mul_ps(qq00,FF),_mm_mul_ps(vftabscale,rinv00)));
 712
 713             /* LENNARD-JONES DISPERSION/REPULSION */
 714
 715             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 716             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 717
 718             fscal            = _mm_add_ps(felec,fvdw);
 719
 720             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 721
 722              /* Update vectorial force */
 723             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 724             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 725             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 726
 727             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 728             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 729             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 730             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 731             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 732                                                    _mm_mul_ps(dx00,fscal),
 733                                                    _mm_mul_ps(dy00,fscal),
 734                                                    _mm_mul_ps(dz00,fscal));
 735
 736             /* Inner loop uses 51 flops */
 737         }
 738
 739         /* End of innermost loop */
 740
 741         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 742                                               f+i_coord_offset,fshift+i_shift_offset);
 743
 744         /* Increment number of inner iterations */
 745         inneriter                  += j_index_end - j_index_start;
 746
 747         /* Outer loop uses 7 flops */
 748     }
 749
 750     /* Increment number of outer iterations */
 751     outeriter        += nri;
 752
 753     /* Update outer/inner flops */
 754
 755     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*51);
 756 }