src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_ElecNone_VdwLJEwSh_GeomP1P1_sse4_1_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 sse4_1_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_sse4_1_single.h"
  48 #include "kernelutil_x86_sse4_1_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwLJEwSh_GeomP1P1_VF_sse4_1_single
  52  * Electrostatics interaction: None
  53  * VdW interaction:            LJEwald
  54  * Geometry:                   Particle-Particle
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecNone_VdwLJEwSh_GeomP1P1_VF_sse4_1_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 SSE, 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           tx,ty,tz,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     int              nvdwtype;
  89     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  90     int              *vdwtype;
  91     real             *vdwparam;
  92     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  93     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  94     __m128           c6grid_00;
  95     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
  96     real             *vdwgridparam;
  97     __m128           one_half  = _mm_set1_ps(0.5);
  98     __m128           minus_one = _mm_set1_ps(-1.0);
  99     __m128           dummy_mask,cutoff_mask;
 100     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 101     __m128           one     = _mm_set1_ps(1.0);
 102     __m128           two     = _mm_set1_ps(2.0);
 103     x                = xx[0];
 104     f                = ff[0];
 105
 106     nri              = nlist->nri;
 107     iinr             = nlist->iinr;
 108     jindex           = nlist->jindex;
 109     jjnr             = nlist->jjnr;
 110     shiftidx         = nlist->shift;
 111     gid              = nlist->gid;
 112     shiftvec         = fr->shift_vec[0];
 113     fshift           = fr->fshift[0];
 114     nvdwtype         = fr->ntype;
 115     vdwparam         = fr->nbfp;
 116     vdwtype          = mdatoms->typeA;
 117     vdwgridparam     = fr->ljpme_c6grid;
 118     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 119     ewclj            = _mm_set1_ps(fr->ewaldcoeff_lj);
 120     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 121
 122     rcutoff_scalar   = fr->rvdw;
 123     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 124     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 125
 126     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 127     rvdw             = _mm_set1_ps(fr->rvdw);
 128
 129     /* Avoid stupid compiler warnings */
 130     jnrA = jnrB = jnrC = jnrD = 0;
 131     j_coord_offsetA = 0;
 132     j_coord_offsetB = 0;
 133     j_coord_offsetC = 0;
 134     j_coord_offsetD = 0;
 135
 136     outeriter        = 0;
 137     inneriter        = 0;
 138
 139     for(iidx=0;iidx<4*DIM;iidx++)
 140     {
 141         scratch[iidx] = 0.0;
 142     }
 143
 144     /* Start outer loop over neighborlists */
 145     for(iidx=0; iidx<nri; iidx++)
 146     {
 147         /* Load shift vector for this list */
 148         i_shift_offset   = DIM*shiftidx[iidx];
 149
 150         /* Load limits for loop over neighbors */
 151         j_index_start    = jindex[iidx];
 152         j_index_end      = jindex[iidx+1];
 153
 154         /* Get outer coordinate index */
 155         inr              = iinr[iidx];
 156         i_coord_offset   = DIM*inr;
 157
 158         /* Load i particle coords and add shift vector */
 159         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 160
 161         fix0             = _mm_setzero_ps();
 162         fiy0             = _mm_setzero_ps();
 163         fiz0             = _mm_setzero_ps();
 164
 165         /* Load parameters for i particles */
 166         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 167
 168         /* Reset potential sums */
 169         vvdwsum          = _mm_setzero_ps();
 170
 171         /* Start inner kernel loop */
 172         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 173         {
 174
 175             /* Get j neighbor index, and coordinate index */
 176             jnrA             = jjnr[jidx];
 177             jnrB             = jjnr[jidx+1];
 178             jnrC             = jjnr[jidx+2];
 179             jnrD             = jjnr[jidx+3];
 180             j_coord_offsetA  = DIM*jnrA;
 181             j_coord_offsetB  = DIM*jnrB;
 182             j_coord_offsetC  = DIM*jnrC;
 183             j_coord_offsetD  = DIM*jnrD;
 184
 185             /* load j atom coordinates */
 186             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 187                                               x+j_coord_offsetC,x+j_coord_offsetD,
 188                                               &jx0,&jy0,&jz0);
 189
 190             /* Calculate displacement vector */
 191             dx00             = _mm_sub_ps(ix0,jx0);
 192             dy00             = _mm_sub_ps(iy0,jy0);
 193             dz00             = _mm_sub_ps(iz0,jz0);
 194
 195             /* Calculate squared distance and things based on it */
 196             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 197
 198             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 199
 200             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 201
 202             /* Load parameters for j particles */
 203             vdwjidx0A        = 2*vdwtype[jnrA+0];
 204             vdwjidx0B        = 2*vdwtype[jnrB+0];
 205             vdwjidx0C        = 2*vdwtype[jnrC+0];
 206             vdwjidx0D        = 2*vdwtype[jnrD+0];
 207
 208             /**************************
 209              * CALCULATE INTERACTIONS *
 210              **************************/
 211
 212             if (gmx_mm_any_lt(rsq00,rcutoff2))
 213             {
 214
 215             r00              = _mm_mul_ps(rsq00,rinv00);
 216
 217             /* Compute parameters for interactions between i and j atoms */
 218             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 219                                          vdwparam+vdwioffset0+vdwjidx0B,
 220                                          vdwparam+vdwioffset0+vdwjidx0C,
 221                                          vdwparam+vdwioffset0+vdwjidx0D,
 222                                          &c6_00,&c12_00);
 223
 224             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 225                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 226                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 227                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 228
 229             /* Analytical LJ-PME */
 230             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 231             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 232             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 233             exponent         = gmx_simd_exp_r(ewcljrsq);
 234             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 235             poly             = _mm_mul_ps(exponent,_mm_add_ps(_mm_sub_ps(one,ewcljrsq),_mm_mul_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half)));
 236             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 237             vvdw6            = _mm_mul_ps(_mm_sub_ps(c6_00,_mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly))),rinvsix);
 238             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 239             vvdw             = _mm_sub_ps(_mm_mul_ps( _mm_sub_ps(vvdw12 , _mm_mul_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6))),one_twelfth),
 240                                _mm_mul_ps( _mm_sub_ps(vvdw6,_mm_add_ps(_mm_mul_ps(c6_00,sh_vdw_invrcut6),_mm_mul_ps(c6grid_00,sh_lj_ewald))),one_sixth));
 241             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 242             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,_mm_sub_ps(vvdw6,_mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6)))),rinvsq00);
 243
 244             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 245
 246             /* Update potential sum for this i atom from the interaction with this j atom. */
 247             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 248             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 249
 250             fscal            = fvdw;
 251
 252             fscal            = _mm_and_ps(fscal,cutoff_mask);
 253
 254             /* Calculate temporary vectorial force */
 255             tx               = _mm_mul_ps(fscal,dx00);
 256             ty               = _mm_mul_ps(fscal,dy00);
 257             tz               = _mm_mul_ps(fscal,dz00);
 258
 259             /* Update vectorial force */
 260             fix0             = _mm_add_ps(fix0,tx);
 261             fiy0             = _mm_add_ps(fiy0,ty);
 262             fiz0             = _mm_add_ps(fiz0,tz);
 263
 264             fjptrA             = f+j_coord_offsetA;
 265             fjptrB             = f+j_coord_offsetB;
 266             fjptrC             = f+j_coord_offsetC;
 267             fjptrD             = f+j_coord_offsetD;
 268             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 269
 270             }
 271
 272             /* Inner loop uses 62 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_add_ps(_mm_sub_ps(one,ewcljrsq),_mm_mul_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half)));
 349             /* vvdw6 = [C6 - C6grid * (1-poly)]/r6 */
 350             vvdw6            = _mm_mul_ps(_mm_sub_ps(c6_00,_mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly))),rinvsix);
 351             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 352             vvdw             = _mm_sub_ps(_mm_mul_ps( _mm_sub_ps(vvdw12 , _mm_mul_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6))),one_twelfth),
 353                                _mm_mul_ps( _mm_sub_ps(vvdw6,_mm_add_ps(_mm_mul_ps(c6_00,sh_vdw_invrcut6),_mm_mul_ps(c6grid_00,sh_lj_ewald))),one_sixth));
 354             /* fvdw = vvdw12/r - (vvdw6/r + (C6grid * exponent * beta^6)/r) */
 355             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,_mm_sub_ps(vvdw6,_mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6)))),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             /* Calculate temporary vectorial force */
 371             tx               = _mm_mul_ps(fscal,dx00);
 372             ty               = _mm_mul_ps(fscal,dy00);
 373             tz               = _mm_mul_ps(fscal,dz00);
 374
 375             /* Update vectorial force */
 376             fix0             = _mm_add_ps(fix0,tx);
 377             fiy0             = _mm_add_ps(fiy0,ty);
 378             fiz0             = _mm_add_ps(fiz0,tz);
 379
 380             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 381             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 382             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 383             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 384             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 385
 386             }
 387
 388             /* Inner loop uses 63 flops */
 389         }
 390
 391         /* End of innermost loop */
 392
 393         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 394                                               f+i_coord_offset,fshift+i_shift_offset);
 395
 396         ggid                        = gid[iidx];
 397         /* Update potential energies */
 398         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 399
 400         /* Increment number of inner iterations */
 401         inneriter                  += j_index_end - j_index_start;
 402
 403         /* Outer loop uses 7 flops */
 404     }
 405
 406     /* Increment number of outer iterations */
 407     outeriter        += nri;
 408
 409     /* Update outer/inner flops */
 410
 411     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_VF,outeriter*7 + inneriter*63);
 412 }
 413 /*
 414  * Gromacs nonbonded kernel:   nb_kernel_ElecNone_VdwLJEwSh_GeomP1P1_F_sse4_1_single
 415  * Electrostatics interaction: None
 416  * VdW interaction:            LJEwald
 417  * Geometry:                   Particle-Particle
 418  * Calculate force/pot:        Force
 419  */
 420 void
 421 nb_kernel_ElecNone_VdwLJEwSh_GeomP1P1_F_sse4_1_single
 422                     (t_nblist                    * gmx_restrict       nlist,
 423                      rvec                        * gmx_restrict          xx,
 424                      rvec                        * gmx_restrict          ff,
 425                      t_forcerec                  * gmx_restrict          fr,
 426                      t_mdatoms                   * gmx_restrict     mdatoms,
 427                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 428                      t_nrnb                      * gmx_restrict        nrnb)
 429 {
 430     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 431      * just 0 for non-waters.
 432      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 433      * jnr indices corresponding to data put in the four positions in the SIMD register.
 434      */
 435     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 436     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 437     int              jnrA,jnrB,jnrC,jnrD;
 438     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 439     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 440     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 441     real             rcutoff_scalar;
 442     real             *shiftvec,*fshift,*x,*f;
 443     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 444     real             scratch[4*DIM];
 445     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 446     int              vdwioffset0;
 447     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 448     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 449     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 450     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 451     int              nvdwtype;
 452     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 453     int              *vdwtype;
 454     real             *vdwparam;
 455     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 456     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 457     __m128           c6grid_00;
 458     __m128           ewclj,ewclj2,ewclj6,ewcljrsq,poly,exponent,f6A,f6B,sh_lj_ewald;
 459     real             *vdwgridparam;
 460     __m128           one_half  = _mm_set1_ps(0.5);
 461     __m128           minus_one = _mm_set1_ps(-1.0);
 462     __m128           dummy_mask,cutoff_mask;
 463     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 464     __m128           one     = _mm_set1_ps(1.0);
 465     __m128           two     = _mm_set1_ps(2.0);
 466     x                = xx[0];
 467     f                = ff[0];
 468
 469     nri              = nlist->nri;
 470     iinr             = nlist->iinr;
 471     jindex           = nlist->jindex;
 472     jjnr             = nlist->jjnr;
 473     shiftidx         = nlist->shift;
 474     gid              = nlist->gid;
 475     shiftvec         = fr->shift_vec[0];
 476     fshift           = fr->fshift[0];
 477     nvdwtype         = fr->ntype;
 478     vdwparam         = fr->nbfp;
 479     vdwtype          = mdatoms->typeA;
 480     vdwgridparam     = fr->ljpme_c6grid;
 481     sh_lj_ewald      = _mm_set1_ps(fr->ic->sh_lj_ewald);
 482     ewclj            = _mm_set1_ps(fr->ewaldcoeff_lj);
 483     ewclj2           = _mm_mul_ps(minus_one,_mm_mul_ps(ewclj,ewclj));
 484
 485     rcutoff_scalar   = fr->rvdw;
 486     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 487     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 488
 489     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 490     rvdw             = _mm_set1_ps(fr->rvdw);
 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         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 530
 531         /* Start inner kernel loop */
 532         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 533         {
 534
 535             /* Get j neighbor index, and coordinate index */
 536             jnrA             = jjnr[jidx];
 537             jnrB             = jjnr[jidx+1];
 538             jnrC             = jjnr[jidx+2];
 539             jnrD             = jjnr[jidx+3];
 540             j_coord_offsetA  = DIM*jnrA;
 541             j_coord_offsetB  = DIM*jnrB;
 542             j_coord_offsetC  = DIM*jnrC;
 543             j_coord_offsetD  = DIM*jnrD;
 544
 545             /* load j atom coordinates */
 546             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 547                                               x+j_coord_offsetC,x+j_coord_offsetD,
 548                                               &jx0,&jy0,&jz0);
 549
 550             /* Calculate displacement vector */
 551             dx00             = _mm_sub_ps(ix0,jx0);
 552             dy00             = _mm_sub_ps(iy0,jy0);
 553             dz00             = _mm_sub_ps(iz0,jz0);
 554
 555             /* Calculate squared distance and things based on it */
 556             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 557
 558             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 559
 560             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 561
 562             /* Load parameters for j particles */
 563             vdwjidx0A        = 2*vdwtype[jnrA+0];
 564             vdwjidx0B        = 2*vdwtype[jnrB+0];
 565             vdwjidx0C        = 2*vdwtype[jnrC+0];
 566             vdwjidx0D        = 2*vdwtype[jnrD+0];
 567
 568             /**************************
 569              * CALCULATE INTERACTIONS *
 570              **************************/
 571
 572             if (gmx_mm_any_lt(rsq00,rcutoff2))
 573             {
 574
 575             r00              = _mm_mul_ps(rsq00,rinv00);
 576
 577             /* Compute parameters for interactions between i and j atoms */
 578             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 579                                          vdwparam+vdwioffset0+vdwjidx0B,
 580                                          vdwparam+vdwioffset0+vdwjidx0C,
 581                                          vdwparam+vdwioffset0+vdwjidx0D,
 582                                          &c6_00,&c12_00);
 583
 584             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 585                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 586                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 587                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 588
 589             /* Analytical LJ-PME */
 590             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 591             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 592             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 593             exponent         = gmx_simd_exp_r(ewcljrsq);
 594             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 595             poly             = _mm_mul_ps(exponent,_mm_add_ps(_mm_sub_ps(one,ewcljrsq),_mm_mul_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half)));
 596             /* f6A = 6 * C6grid * (1 - poly) */
 597             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
 598             /* f6B = C6grid * exponent * beta^6 */
 599             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
 600             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 601             fvdw              = _mm_mul_ps(_mm_add_ps(_mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),_mm_sub_ps(c6_00,f6A)),rinvsix),f6B),rinvsq00);
 602
 603             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 604
 605             fscal            = fvdw;
 606
 607             fscal            = _mm_and_ps(fscal,cutoff_mask);
 608
 609             /* Calculate temporary vectorial force */
 610             tx               = _mm_mul_ps(fscal,dx00);
 611             ty               = _mm_mul_ps(fscal,dy00);
 612             tz               = _mm_mul_ps(fscal,dz00);
 613
 614             /* Update vectorial force */
 615             fix0             = _mm_add_ps(fix0,tx);
 616             fiy0             = _mm_add_ps(fiy0,ty);
 617             fiz0             = _mm_add_ps(fiz0,tz);
 618
 619             fjptrA             = f+j_coord_offsetA;
 620             fjptrB             = f+j_coord_offsetB;
 621             fjptrC             = f+j_coord_offsetC;
 622             fjptrD             = f+j_coord_offsetD;
 623             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 624
 625             }
 626
 627             /* Inner loop uses 49 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             vdwjidx0A        = 2*vdwtype[jnrA+0];
 671             vdwjidx0B        = 2*vdwtype[jnrB+0];
 672             vdwjidx0C        = 2*vdwtype[jnrC+0];
 673             vdwjidx0D        = 2*vdwtype[jnrD+0];
 674
 675             /**************************
 676              * CALCULATE INTERACTIONS *
 677              **************************/
 678
 679             if (gmx_mm_any_lt(rsq00,rcutoff2))
 680             {
 681
 682             r00              = _mm_mul_ps(rsq00,rinv00);
 683             r00              = _mm_andnot_ps(dummy_mask,r00);
 684
 685             /* Compute parameters for interactions between i and j atoms */
 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             c6grid_00       = gmx_mm_load_4real_swizzle_ps(vdwgridparam+vdwioffset0+vdwjidx0A,
 693                                                                vdwgridparam+vdwioffset0+vdwjidx0B,
 694                                                                vdwgridparam+vdwioffset0+vdwjidx0C,
 695                                                                vdwgridparam+vdwioffset0+vdwjidx0D);
 696
 697             /* Analytical LJ-PME */
 698             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 699             ewcljrsq         = _mm_mul_ps(ewclj2,rsq00);
 700             ewclj6           = _mm_mul_ps(ewclj2,_mm_mul_ps(ewclj2,ewclj2));
 701             exponent         = gmx_simd_exp_r(ewcljrsq);
 702             /* poly = exp(-(beta*r)^2) * (1 + (beta*r)^2 + (beta*r)^4 /2) */
 703             poly             = _mm_mul_ps(exponent,_mm_add_ps(_mm_sub_ps(one,ewcljrsq),_mm_mul_ps(_mm_mul_ps(ewcljrsq,ewcljrsq),one_half)));
 704             /* f6A = 6 * C6grid * (1 - poly) */
 705             f6A              = _mm_mul_ps(c6grid_00,_mm_sub_ps(one,poly));
 706             /* f6B = C6grid * exponent * beta^6 */
 707             f6B              = _mm_mul_ps(_mm_mul_ps(c6grid_00,one_sixth),_mm_mul_ps(exponent,ewclj6));
 708             /* fvdw = 12*C12/r13 - ((6*C6 - f6A)/r6 + f6B)/r */
 709             fvdw              = _mm_mul_ps(_mm_add_ps(_mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),_mm_sub_ps(c6_00,f6A)),rinvsix),f6B),rinvsq00);
 710
 711             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 712
 713             fscal            = fvdw;
 714
 715             fscal            = _mm_and_ps(fscal,cutoff_mask);
 716
 717             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 718
 719             /* Calculate temporary vectorial force */
 720             tx               = _mm_mul_ps(fscal,dx00);
 721             ty               = _mm_mul_ps(fscal,dy00);
 722             tz               = _mm_mul_ps(fscal,dz00);
 723
 724             /* Update vectorial force */
 725             fix0             = _mm_add_ps(fix0,tx);
 726             fiy0             = _mm_add_ps(fiy0,ty);
 727             fiz0             = _mm_add_ps(fiz0,tz);
 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,tx,ty,tz);
 734
 735             }
 736
 737             /* Inner loop uses 50 flops */
 738         }
 739
 740         /* End of innermost loop */
 741
 742         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 743                                               f+i_coord_offset,fshift+i_shift_offset);
 744
 745         /* Increment number of inner iterations */
 746         inneriter                  += j_index_end - j_index_start;
 747
 748         /* Outer loop uses 6 flops */
 749     }
 750
 751     /* Increment number of outer iterations */
 752     outeriter        += nri;
 753
 754     /* Update outer/inner flops */
 755
 756     inc_nrnb(nrnb,eNR_NBKERNEL_VDW_F,outeriter*6 + inneriter*50);
 757 }