src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecNone_VdwLJEwSh_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_ElecNone_VdwLJEwSh_GeomP1P1_VF_avx_128_fma_single
  54  * Electrostatics interaction: None
  55  * VdW interaction:            LJEwald
  56  * Geometry:                   Particle-Particle
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecNone_VdwLJEwSh_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     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     __m128           c6grid_00;
  97     real             *vdwgridparam;
  98     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
  99     __m128           one_half = _mm_set1_ps(0.5);
 100     __m128           minus_one = _mm_set1_ps(-1.0);
 101     __m128           dummy_mask,cutoff_mask;
 102     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 103     __m128           one     = _mm_set1_ps(1.0);
 104     __m128           two     = _mm_set1_ps(2.0);
 105     x                = xx[0];
 106     f                = ff[0];
 107
 108     nri              = nlist->nri;
 109     iinr             = nlist->iinr;
 110     jindex           = nlist->jindex;
 111     jjnr             = nlist->jjnr;
 112     shiftidx         = nlist->shift;
 113     gid              = nlist->gid;
 114     shiftvec         = fr->shift_vec[0];
 115     fshift           = fr->fshift[0];
 116     nvdwtype         = fr->ntype;
 117     vdwparam         = fr->nbfp;
 118     vdwtype          = mdatoms->typeA;
 119     vdwgridparam     = fr->ljpme_c6grid;
 120     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 121     ewclj            = _mm_set1_ps(fr->ewaldcoeff_lj);
 122     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 123
 124     rcutoff_scalar   = fr->rvdw;
 125     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 126     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 127
 128     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 129     rvdw             = _mm_set1_ps(fr->rvdw);
 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         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 169
 170         /* Reset potential sums */
 171         vvdwsum          = _mm_setzero_ps();
 172
 173         /* Start inner kernel loop */
 174         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 175         {
 176
 177             /* Get j neighbor index, and coordinate index */
 178             jnrA             = jjnr[jidx];
 179             jnrB             = jjnr[jidx+1];
 180             jnrC             = jjnr[jidx+2];
 181             jnrD             = jjnr[jidx+3];
 182             j_coord_offsetA  = DIM*jnrA;
 183             j_coord_offsetB  = DIM*jnrB;
 184             j_coord_offsetC  = DIM*jnrC;
 185             j_coord_offsetD  = DIM*jnrD;
 186
 187             /* load j atom coordinates */
 188             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 189                                               x+j_coord_offsetC,x+j_coord_offsetD,
 190                                               &jx0,&jy0,&jz0);
 191
 192             /* Calculate displacement vector */
 193             dx00             = _mm_sub_ps(ix0,jx0);
 194             dy00             = _mm_sub_ps(iy0,jy0);
 195             dz00             = _mm_sub_ps(iz0,jz0);
 196
 197             /* Calculate squared distance and things based on it */
 198             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 199
 200             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 201
 202             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 203
 204             /* Load parameters for j particles */
 205             vdwjidx0A        = 2*vdwtype[jnrA+0];
 206             vdwjidx0B        = 2*vdwtype[jnrB+0];
 207             vdwjidx0C        = 2*vdwtype[jnrC+0];
 208             vdwjidx0D        = 2*vdwtype[jnrD+0];
 209
 210             /**************************
 211              * CALCULATE INTERACTIONS *
 212              **************************/
 213
 214             if (gmx_mm_any_lt(rsq00,rcutoff2))
 215             {
 216
 217             r00              = _mm_mul_ps(rsq00,rinv00);
 218
 219             /* Compute parameters for interactions between i and j atoms */
 220             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 221                                          vdwparam+vdwioffset0+vdwjidx0B,
 222                                          vdwparam+vdwioffset0+vdwjidx0C,
 223                                          vdwparam+vdwioffset0+vdwjidx0D,
 224                                          &c6_00,&c12_00);
 225
 226             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 227                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 228                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 229                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 230
 231             /* Analytical LJ-PME */
 232             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 233             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 234             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 235             exponent         = gmx_simd_exp_r(ewcljrsq);
 236             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 237             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 238             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 239             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_00,_mm_sub_ps(one,poly),c6_00),rinvsix);
 240             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 241             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 242                                           _mm_mul_ps(_mm_sub_ps(vvdw6,_mm_macc_ps(c6grid_00,sh_lj_ewald,_mm_mul_ps(c6_00,sh_vdw_invrcut6))),one_sixth));
 243             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 244             fvdw             = _mm_mul_ps(_mm_add_ps(vvdw12,_mm_msub_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6),vvdw6)),rinvsq00);
 245
 246             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 247
 248             /* Update potential sum for this i atom from the interaction with this j atom. */
 249             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 250             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 251
 252             fscal            = fvdw;
 253
 254             fscal            = _mm_and_ps(fscal,cutoff_mask);
 255
 256              /* Update vectorial force */
 257             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 258             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 259             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 260
 261             fjptrA             = f+j_coord_offsetA;
 262             fjptrB             = f+j_coord_offsetB;
 263             fjptrC             = f+j_coord_offsetC;
 264             fjptrD             = f+j_coord_offsetD;
 265             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 266                                                    _mm_mul_ps(dx00,fscal),
 267                                                    _mm_mul_ps(dy00,fscal),
 268                                                    _mm_mul_ps(dz00,fscal));
 269
 270             }
 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             vdwjidx0A        = 2*vdwtype[jnrA+0];
 316             vdwjidx0B        = 2*vdwtype[jnrB+0];
 317             vdwjidx0C        = 2*vdwtype[jnrC+0];
 318             vdwjidx0D        = 2*vdwtype[jnrD+0];
 319
 320             /**************************
 321              * CALCULATE INTERACTIONS *
 322              **************************/
 323
 324             if (gmx_mm_any_lt(rsq00,rcutoff2))
 325             {
 326
 327             r00              = _mm_mul_ps(rsq00,rinv00);
 328             r00              = _mm_andnot_ps(dummy_mask,r00);
 329
 330             /* Compute parameters for interactions between i and j atoms */
 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             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 338                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 339                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 340                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 341
 342             /* Analytical LJ-PME */
 343             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 344             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 345             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 346             exponent         = gmx_simd_exp_r(ewcljrsq);
 347             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 348             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 349             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 350             vvdw6            = _mm_mul_ps(_mm_macc_ps(-c6grid_00,_mm_sub_ps(one,poly),c6_00),rinvsix);
 351             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 352             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 353                                           _mm_mul_ps(_mm_sub_ps(vvdw6,_mm_macc_ps(c6grid_00,sh_lj_ewald,_mm_mul_ps(c6_00,sh_vdw_invrcut6))),one_sixth));
 354             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 355             fvdw             = _mm_mul_ps(_mm_add_ps(vvdw12,_mm_msub_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6),vvdw6)),rinvsq00);
 356
 357             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 358
 359             /* Update potential sum for this i atom from the interaction with this j atom. */
 360             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 361             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 362             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 363
 364             fscal            = fvdw;
 365
 366             fscal            = _mm_and_ps(fscal,cutoff_mask);
 367
 368             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 369
 370              /* Update vectorial force */
 371             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 372             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 373             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 374
 375             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 376             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 377             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 378             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 379             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 380                                                    _mm_mul_ps(dx00,fscal),
 381                                                    _mm_mul_ps(dy00,fscal),
 382                                                    _mm_mul_ps(dz00,fscal));
 383
 384             }
 385
 386             /* Inner loop uses 60 flops */
 387         }
 388
 389         /* End of innermost loop */
 390
 391         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 392                                               f+i_coord_offset,fshift+i_shift_offset);
 393
 394         ggid                        = gid[iidx];
 395         /* Update potential energies */
 396         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 397
 398         /* Increment number of inner iterations */
 399         inneriter                  += j_index_end - j_index_start;
 400
 401         /* Outer loop uses 7 flops */
 402     }
 403
 404     /* Increment number of outer iterations */
 405     outeriter        += nri;
 406
 407     /* Update outer/inner flops */
 408
 409     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_VF,outeriter*7 + inneriter*60);
 410 }
 411 /*
 412  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwLJEwSh_GeomP1P1_F_avx_128_fma_single
 413  * Electrostatics interaction: None
 414  * VdW interaction:            LJEwald
 415  * Geometry:                   Particle-Particle
 416  * Calculate force/pot:        Force
 417  */
 418 void
 419 nb_kernel_ElecNone_VdwLJEwSh_GeomP1P1_F_avx_128_fma_single
 420                     (t_nblist                    * gmx_restrict       nlist,
 421                      rvec                        * gmx_restrict          xx,
 422                      rvec                        * gmx_restrict          ff,
 423                      t_forcerec                  * gmx_restrict          fr,
 424                      t_mdatoms                   * gmx_restrict     mdatoms,
 425                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 426                      t_nrnb                      * gmx_restrict        nrnb)
 427 {
 428     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 429      * just 0 for non-waters.
 430      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 431      * jnr indices corresponding to data put in the four positions in the SIMD register.
 432      */
 433     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 434     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 435     int              jnrA,jnrB,jnrC,jnrD;
 436     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 437     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 438     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 439     real             rcutoff_scalar;
 440     real             *shiftvec,*fshift,*x,*f;
 441     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 442     real             scratch[4*DIM];
 443     __m128           fscal,rcutoff,rcutoff2,jidxall;
 444     int              vdwioffset0;
 445     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 446     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 447     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 448     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 449     int              nvdwtype;
 450     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 451     int              *vdwtype;
 452     real             *vdwparam;
 453     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 454     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 455     __m128           c6grid_00;
 456     real             *vdwgridparam;
 457     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 458     __m128           one_half = _mm_set1_ps(0.5);
 459     __m128           minus_one = _mm_set1_ps(-1.0);
 460     __m128           dummy_mask,cutoff_mask;
 461     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 462     __m128           one     = _mm_set1_ps(1.0);
 463     __m128           two     = _mm_set1_ps(2.0);
 464     x                = xx[0];
 465     f                = ff[0];
 466
 467     nri              = nlist->nri;
 468     iinr             = nlist->iinr;
 469     jindex           = nlist->jindex;
 470     jjnr             = nlist->jjnr;
 471     shiftidx         = nlist->shift;
 472     gid              = nlist->gid;
 473     shiftvec         = fr->shift_vec[0];
 474     fshift           = fr->fshift[0];
 475     nvdwtype         = fr->ntype;
 476     vdwparam         = fr->nbfp;
 477     vdwtype          = mdatoms->typeA;
 478     vdwgridparam     = fr->ljpme_c6grid;
 479     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 480     ewclj            = _mm_set1_ps(fr->ewaldcoeff_lj);
 481     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 482
 483     rcutoff_scalar   = fr->rvdw;
 484     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 485     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 486
 487     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 488     rvdw             = _mm_set1_ps(fr->rvdw);
 489
 490     /* Avoid stupid compiler warnings */
 491     jnrA = jnrB = jnrC = jnrD = 0;
 492     j_coord_offsetA = 0;
 493     j_coord_offsetB = 0;
 494     j_coord_offsetC = 0;
 495     j_coord_offsetD = 0;
 496
 497     outeriter        = 0;
 498     inneriter        = 0;
 499
 500     for(iidx=0;iidx<4*DIM;iidx++)
 501     {
 502         scratch[iidx] = 0.0;
 503     }
 504
 505     /* Start outer loop over neighborlists */
 506     for(iidx=0; iidx<nri; iidx++)
 507     {
 508         /* Load shift vector for this list */
 509         i_shift_offset   = DIM*shiftidx[iidx];
 510
 511         /* Load limits for loop over neighbors */
 512         j_index_start    = jindex[iidx];
 513         j_index_end      = jindex[iidx+1];
 514
 515         /* Get outer coordinate index */
 516         inr              = iinr[iidx];
 517         i_coord_offset   = DIM*inr;
 518
 519         /* Load i particle coords and add shift vector */
 520         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 521
 522         fix0             = _mm_setzero_ps();
 523         fiy0             = _mm_setzero_ps();
 524         fiz0             = _mm_setzero_ps();
 525
 526         /* Load parameters for i particles */
 527         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 528
 529         /* Start inner kernel loop */
 530         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 531         {
 532
 533             /* Get j neighbor index, and coordinate index */
 534             jnrA             = jjnr[jidx];
 535             jnrB             = jjnr[jidx+1];
 536             jnrC             = jjnr[jidx+2];
 537             jnrD             = jjnr[jidx+3];
 538             j_coord_offsetA  = DIM*jnrA;
 539             j_coord_offsetB  = DIM*jnrB;
 540             j_coord_offsetC  = DIM*jnrC;
 541             j_coord_offsetD  = DIM*jnrD;
 542
 543             /* load j atom coordinates */
 544             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 545                                               x+j_coord_offsetC,x+j_coord_offsetD,
 546                                               &jx0,&jy0,&jz0);
 547
 548             /* Calculate displacement vector */
 549             dx00             = _mm_sub_ps(ix0,jx0);
 550             dy00             = _mm_sub_ps(iy0,jy0);
 551             dz00             = _mm_sub_ps(iz0,jz0);
 552
 553             /* Calculate squared distance and things based on it */
 554             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 555
 556             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 557
 558             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 559
 560             /* Load parameters for j particles */
 561             vdwjidx0A        = 2*vdwtype[jnrA+0];
 562             vdwjidx0B        = 2*vdwtype[jnrB+0];
 563             vdwjidx0C        = 2*vdwtype[jnrC+0];
 564             vdwjidx0D        = 2*vdwtype[jnrD+0];
 565
 566             /**************************
 567              * CALCULATE INTERACTIONS *
 568              **************************/
 569
 570             if (gmx_mm_any_lt(rsq00,rcutoff2))
 571             {
 572
 573             r00              = _mm_mul_ps(rsq00,rinv00);
 574
 575             /* Compute parameters for interactions between i and j atoms */
 576             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 577                                          vdwparam+vdwioffset0+vdwjidx0B,
 578                                          vdwparam+vdwioffset0+vdwjidx0C,
 579                                          vdwparam+vdwioffset0+vdwjidx0D,
 580                                          &c6_00,&c12_00);
 581
 582             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 583                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 584                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 585                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 586
 587             /* Analytical LJ-PME */
 588             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 589             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 590             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 591             exponent         = gmx_simd_exp_r(ewcljrsq);
 592             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 593             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 594             /* f6A = 6 * C6grid * (1 - poly) */
 595             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
 596             /* f6B = C6grid * exponent * beta^6 */
 597             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
 598             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 599             fvdw              = _mm_mul_ps(_mm_macc_ps(_mm_msub_ps(c12_00,rinvsix,_mm_sub_ps(c6_00,f6A)),rinvsix,f6B),rinvsq00);
 600
 601             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 602
 603             fscal            = fvdw;
 604
 605             fscal            = _mm_and_ps(fscal,cutoff_mask);
 606
 607              /* Update vectorial force */
 608             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 609             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 610             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 611
 612             fjptrA             = f+j_coord_offsetA;
 613             fjptrB             = f+j_coord_offsetB;
 614             fjptrC             = f+j_coord_offsetC;
 615             fjptrD             = f+j_coord_offsetD;
 616             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 617                                                    _mm_mul_ps(dx00,fscal),
 618                                                    _mm_mul_ps(dy00,fscal),
 619                                                    _mm_mul_ps(dz00,fscal));
 620
 621             }
 622
 623             /* Inner loop uses 50 flops */
 624         }
 625
 626         if(jidx<j_index_end)
 627         {
 628
 629             /* Get j neighbor index, and coordinate index */
 630             jnrlistA         = jjnr[jidx];
 631             jnrlistB         = jjnr[jidx+1];
 632             jnrlistC         = jjnr[jidx+2];
 633             jnrlistD         = jjnr[jidx+3];
 634             /* Sign of each element will be negative for non-real atoms.
 635              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 636              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 637              */
 638             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 639             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 640             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 641             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 642             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 643             j_coord_offsetA  = DIM*jnrA;
 644             j_coord_offsetB  = DIM*jnrB;
 645             j_coord_offsetC  = DIM*jnrC;
 646             j_coord_offsetD  = DIM*jnrD;
 647
 648             /* load j atom coordinates */
 649             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 650                                               x+j_coord_offsetC,x+j_coord_offsetD,
 651                                               &jx0,&jy0,&jz0);
 652
 653             /* Calculate displacement vector */
 654             dx00             = _mm_sub_ps(ix0,jx0);
 655             dy00             = _mm_sub_ps(iy0,jy0);
 656             dz00             = _mm_sub_ps(iz0,jz0);
 657
 658             /* Calculate squared distance and things based on it */
 659             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 660
 661             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 662
 663             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 664
 665             /* Load parameters for j particles */
 666             vdwjidx0A        = 2*vdwtype[jnrA+0];
 667             vdwjidx0B        = 2*vdwtype[jnrB+0];
 668             vdwjidx0C        = 2*vdwtype[jnrC+0];
 669             vdwjidx0D        = 2*vdwtype[jnrD+0];
 670
 671             /**************************
 672              * CALCULATE INTERACTIONS *
 673              **************************/
 674
 675             if (gmx_mm_any_lt(rsq00,rcutoff2))
 676             {
 677
 678             r00              = _mm_mul_ps(rsq00,rinv00);
 679             r00              = _mm_andnot_ps(dummy_mask,r00);
 680
 681             /* Compute parameters for interactions between i and j atoms */
 682             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 683                                          vdwparam+vdwioffset0+vdwjidx0B,
 684                                          vdwparam+vdwioffset0+vdwjidx0C,
 685                                          vdwparam+vdwioffset0+vdwjidx0D,
 686                                          &c6_00,&c12_00);
 687
 688             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 689                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 690                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 691                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 692
 693             /* Analytical LJ-PME */
 694             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 695             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 696             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 697             exponent         = gmx_simd_exp_r(ewcljrsq);
 698             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 699             poly             = _mm_mul_ps(exponent,_mm_macc_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half,_mm_sub_ps(one,ewcljrsq)));
 700             /* f6A = 6 * C6grid * (1 - poly) */
 701             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
 702             /* f6B = C6grid * exponent * beta^6 */
 703             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
 704             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 705             fvdw              = _mm_mul_ps(_mm_macc_ps(_mm_msub_ps(c12_00,rinvsix,_mm_sub_ps(c6_00,f6A)),rinvsix,f6B),rinvsq00);
 706
 707             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 708
 709             fscal            = fvdw;
 710
 711             fscal            = _mm_and_ps(fscal,cutoff_mask);
 712
 713             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 714
 715              /* Update vectorial force */
 716             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 717             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 718             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 719
 720             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 721             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 722             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 723             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 724             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 725                                                    _mm_mul_ps(dx00,fscal),
 726                                                    _mm_mul_ps(dy00,fscal),
 727                                                    _mm_mul_ps(dz00,fscal));
 728
 729             }
 730
 731             /* Inner loop uses 51 flops */
 732         }
 733
 734         /* End of innermost loop */
 735
 736         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 737                                               f+i_coord_offset,fshift+i_shift_offset);
 738
 739         /* Increment number of inner iterations */
 740         inneriter                  += j_index_end - j_index_start;
 741
 742         /* Outer loop uses 6 flops */
 743     }
 744
 745     /* Increment number of outer iterations */
 746     outeriter        += nri;
 747
 748     /* Update outer/inner flops */
 749
 750     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_F,outeriter*6 + inneriter*51);
 751 }