src/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_ElecEwSh_VdwLJSh_GeomP1P1_sse4_1_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse4_1_single kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_sse4_1_single.h"
  34 #include "kernelutil_x86_sse4_1_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJSh_GeomP1P1_VF_sse4_1_single
  38  * Electrostatics interaction: Ewald
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecEwSh_VdwLJSh_GeomP1P1_VF_sse4_1_single
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB,jnrC,jnrD;
  61     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  62     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  63     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  64     real             rcutoff_scalar;
  65     real             *shiftvec,*fshift,*x,*f;
  66     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  67     real             scratch[4*DIM];
  68     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  69     int              vdwioffset0;
  70     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  71     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  72     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  73     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  74     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
  75     real             *charge;
  76     int              nvdwtype;
  77     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  78     int              *vdwtype;
  79     real             *vdwparam;
  80     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
  81     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
  82     __m128i          ewitab;
  83     __m128           ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  84     real             *ewtab;
  85     __m128           dummy_mask,cutoff_mask;
  86     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  87     __m128           one     = _mm_set1_ps(1.0);
  88     __m128           two     = _mm_set1_ps(2.0);
  89     x                = xx[0];
  90     f                = ff[0];
  91
  92     nri              = nlist->nri;
  93     iinr             = nlist->iinr;
  94     jindex           = nlist->jindex;
  95     jjnr             = nlist->jjnr;
  96     shiftidx         = nlist->shift;
  97     gid              = nlist->gid;
  98     shiftvec         = fr->shift_vec[0];
  99     fshift           = fr->fshift[0];
 100     facel            = _mm_set1_ps(fr->epsfac);
 101     charge           = mdatoms->chargeA;
 102     nvdwtype         = fr->ntype;
 103     vdwparam         = fr->nbfp;
 104     vdwtype          = mdatoms->typeA;
 105
 106     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 107     ewtab            = fr->ic->tabq_coul_FDV0;
 108     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 109     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 110
 111     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 112     rcutoff_scalar   = fr->rcoulomb;
 113     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 114     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 115
 116     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 117     rvdw             = _mm_set1_ps(fr->rvdw);
 118
 119     /* Avoid stupid compiler warnings */
 120     jnrA = jnrB = jnrC = jnrD = 0;
 121     j_coord_offsetA = 0;
 122     j_coord_offsetB = 0;
 123     j_coord_offsetC = 0;
 124     j_coord_offsetD = 0;
 125
 126     outeriter        = 0;
 127     inneriter        = 0;
 128
 129     for(iidx=0;iidx<4*DIM;iidx++)
 130     {
 131         scratch[iidx] = 0.0;
 132     }
 133
 134     /* Start outer loop over neighborlists */
 135     for(iidx=0; iidx<nri; iidx++)
 136     {
 137         /* Load shift vector for this list */
 138         i_shift_offset   = DIM*shiftidx[iidx];
 139
 140         /* Load limits for loop over neighbors */
 141         j_index_start    = jindex[iidx];
 142         j_index_end      = jindex[iidx+1];
 143
 144         /* Get outer coordinate index */
 145         inr              = iinr[iidx];
 146         i_coord_offset   = DIM*inr;
 147
 148         /* Load i particle coords and add shift vector */
 149         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 150
 151         fix0             = _mm_setzero_ps();
 152         fiy0             = _mm_setzero_ps();
 153         fiz0             = _mm_setzero_ps();
 154
 155         /* Load parameters for i particles */
 156         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 157         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 158
 159         /* Reset potential sums */
 160         velecsum         = _mm_setzero_ps();
 161         vvdwsum          = _mm_setzero_ps();
 162
 163         /* Start inner kernel loop */
 164         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 165         {
 166
 167             /* Get j neighbor index, and coordinate index */
 168             jnrA             = jjnr[jidx];
 169             jnrB             = jjnr[jidx+1];
 170             jnrC             = jjnr[jidx+2];
 171             jnrD             = jjnr[jidx+3];
 172             j_coord_offsetA  = DIM*jnrA;
 173             j_coord_offsetB  = DIM*jnrB;
 174             j_coord_offsetC  = DIM*jnrC;
 175             j_coord_offsetD  = DIM*jnrD;
 176
 177             /* load j atom coordinates */
 178             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 179                                               x+j_coord_offsetC,x+j_coord_offsetD,
 180                                               &jx0,&jy0,&jz0);
 181
 182             /* Calculate displacement vector */
 183             dx00             = _mm_sub_ps(ix0,jx0);
 184             dy00             = _mm_sub_ps(iy0,jy0);
 185             dz00             = _mm_sub_ps(iz0,jz0);
 186
 187             /* Calculate squared distance and things based on it */
 188             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 189
 190             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 191
 192             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 193
 194             /* Load parameters for j particles */
 195             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 196                                                               charge+jnrC+0,charge+jnrD+0);
 197             vdwjidx0A        = 2*vdwtype[jnrA+0];
 198             vdwjidx0B        = 2*vdwtype[jnrB+0];
 199             vdwjidx0C        = 2*vdwtype[jnrC+0];
 200             vdwjidx0D        = 2*vdwtype[jnrD+0];
 201
 202             /**************************
 203              * CALCULATE INTERACTIONS *
 204              **************************/
 205
 206             if (gmx_mm_any_lt(rsq00,rcutoff2))
 207             {
 208
 209             r00              = _mm_mul_ps(rsq00,rinv00);
 210
 211             /* Compute parameters for interactions between i and j atoms */
 212             qq00             = _mm_mul_ps(iq0,jq0);
 213             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 214                                          vdwparam+vdwioffset0+vdwjidx0B,
 215                                          vdwparam+vdwioffset0+vdwjidx0C,
 216                                          vdwparam+vdwioffset0+vdwjidx0D,
 217                                          &c6_00,&c12_00);
 218
 219             /* EWALD ELECTROSTATICS */
 220
 221             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 222             ewrt             = _mm_mul_ps(r00,ewtabscale);
 223             ewitab           = _mm_cvttps_epi32(ewrt);
 224             eweps            = _mm_sub_ps(ewrt,_mm_round_ps(ewrt, _MM_FROUND_FLOOR));
 225             ewitab           = _mm_slli_epi32(ewitab,2);
 226             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 227             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 228             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 229             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 230             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 231             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 232             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 233             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_sub_ps(rinv00,sh_ewald),velec));
 234             felec            = _mm_mul_ps(_mm_mul_ps(qq00,rinv00),_mm_sub_ps(rinvsq00,felec));
 235
 236             /* LENNARD-JONES DISPERSION/REPULSION */
 237
 238             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 239             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 240             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 241             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) ,
 242                                           _mm_mul_ps( _mm_sub_ps(vvdw6,_mm_mul_ps(c6_00,sh_vdw_invrcut6)),one_sixth));
 243             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 244
 245             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 246
 247             /* Update potential sum for this i atom from the interaction with this j atom. */
 248             velec            = _mm_and_ps(velec,cutoff_mask);
 249             velecsum         = _mm_add_ps(velecsum,velec);
 250             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 251             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 252
 253             fscal            = _mm_add_ps(felec,fvdw);
 254
 255             fscal            = _mm_and_ps(fscal,cutoff_mask);
 256
 257             /* Calculate temporary vectorial force */
 258             tx               = _mm_mul_ps(fscal,dx00);
 259             ty               = _mm_mul_ps(fscal,dy00);
 260             tz               = _mm_mul_ps(fscal,dz00);
 261
 262             /* Update vectorial force */
 263             fix0             = _mm_add_ps(fix0,tx);
 264             fiy0             = _mm_add_ps(fiy0,ty);
 265             fiz0             = _mm_add_ps(fiz0,tz);
 266
 267             fjptrA             = f+j_coord_offsetA;
 268             fjptrB             = f+j_coord_offsetB;
 269             fjptrC             = f+j_coord_offsetC;
 270             fjptrD             = f+j_coord_offsetD;
 271             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 272
 273             }
 274
 275             /* Inner loop uses 64 flops */
 276         }
 277
 278         if(jidx<j_index_end)
 279         {
 280
 281             /* Get j neighbor index, and coordinate index */
 282             jnrlistA         = jjnr[jidx];
 283             jnrlistB         = jjnr[jidx+1];
 284             jnrlistC         = jjnr[jidx+2];
 285             jnrlistD         = jjnr[jidx+3];
 286             /* Sign of each element will be negative for non-real atoms.
 287              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 288              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 289              */
 290             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 291             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 292             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 293             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 294             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 295             j_coord_offsetA  = DIM*jnrA;
 296             j_coord_offsetB  = DIM*jnrB;
 297             j_coord_offsetC  = DIM*jnrC;
 298             j_coord_offsetD  = DIM*jnrD;
 299
 300             /* load j atom coordinates */
 301             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 302                                               x+j_coord_offsetC,x+j_coord_offsetD,
 303                                               &jx0,&jy0,&jz0);
 304
 305             /* Calculate displacement vector */
 306             dx00             = _mm_sub_ps(ix0,jx0);
 307             dy00             = _mm_sub_ps(iy0,jy0);
 308             dz00             = _mm_sub_ps(iz0,jz0);
 309
 310             /* Calculate squared distance and things based on it */
 311             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 312
 313             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 314
 315             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 316
 317             /* Load parameters for j particles */
 318             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 319                                                               charge+jnrC+0,charge+jnrD+0);
 320             vdwjidx0A        = 2*vdwtype[jnrA+0];
 321             vdwjidx0B        = 2*vdwtype[jnrB+0];
 322             vdwjidx0C        = 2*vdwtype[jnrC+0];
 323             vdwjidx0D        = 2*vdwtype[jnrD+0];
 324
 325             /**************************
 326              * CALCULATE INTERACTIONS *
 327              **************************/
 328
 329             if (gmx_mm_any_lt(rsq00,rcutoff2))
 330             {
 331
 332             r00              = _mm_mul_ps(rsq00,rinv00);
 333             r00              = _mm_andnot_ps(dummy_mask,r00);
 334
 335             /* Compute parameters for interactions between i and j atoms */
 336             qq00             = _mm_mul_ps(iq0,jq0);
 337             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 338                                          vdwparam+vdwioffset0+vdwjidx0B,
 339                                          vdwparam+vdwioffset0+vdwjidx0C,
 340                                          vdwparam+vdwioffset0+vdwjidx0D,
 341                                          &c6_00,&c12_00);
 342
 343             /* EWALD ELECTROSTATICS */
 344
 345             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 346             ewrt             = _mm_mul_ps(r00,ewtabscale);
 347             ewitab           = _mm_cvttps_epi32(ewrt);
 348             eweps            = _mm_sub_ps(ewrt,_mm_round_ps(ewrt, _MM_FROUND_FLOOR));
 349             ewitab           = _mm_slli_epi32(ewitab,2);
 350             ewtabF           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 351             ewtabD           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 352             ewtabV           = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,2) );
 353             ewtabFn          = _mm_load_ps( ewtab + gmx_mm_extract_epi32(ewitab,3) );
 354             _MM_TRANSPOSE4_PS(ewtabF,ewtabD,ewtabV,ewtabFn);
 355             felec            = _mm_add_ps(ewtabF,_mm_mul_ps(eweps,ewtabD));
 356             velec            = _mm_sub_ps(ewtabV,_mm_mul_ps(_mm_mul_ps(ewtabhalfspace,eweps),_mm_add_ps(ewtabF,felec)));
 357             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_sub_ps(rinv00,sh_ewald),velec));
 358             felec            = _mm_mul_ps(_mm_mul_ps(qq00,rinv00),_mm_sub_ps(rinvsq00,felec));
 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_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) ,
 366                                           _mm_mul_ps( _mm_sub_ps(vvdw6,_mm_mul_ps(c6_00,sh_vdw_invrcut6)),one_sixth));
 367             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 368
 369             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 370
 371             /* Update potential sum for this i atom from the interaction with this j atom. */
 372             velec            = _mm_and_ps(velec,cutoff_mask);
 373             velec            = _mm_andnot_ps(dummy_mask,velec);
 374             velecsum         = _mm_add_ps(velecsum,velec);
 375             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 376             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 377             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 378
 379             fscal            = _mm_add_ps(felec,fvdw);
 380
 381             fscal            = _mm_and_ps(fscal,cutoff_mask);
 382
 383             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 384
 385             /* Calculate temporary vectorial force */
 386             tx               = _mm_mul_ps(fscal,dx00);
 387             ty               = _mm_mul_ps(fscal,dy00);
 388             tz               = _mm_mul_ps(fscal,dz00);
 389
 390             /* Update vectorial force */
 391             fix0             = _mm_add_ps(fix0,tx);
 392             fiy0             = _mm_add_ps(fiy0,ty);
 393             fiz0             = _mm_add_ps(fiz0,tz);
 394
 395             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 396             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 397             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 398             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 399             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 400
 401             }
 402
 403             /* Inner loop uses 65 flops */
 404         }
 405
 406         /* End of innermost loop */
 407
 408         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 409                                               f+i_coord_offset,fshift+i_shift_offset);
 410
 411         ggid                        = gid[iidx];
 412         /* Update potential energies */
 413         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 414         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 415
 416         /* Increment number of inner iterations */
 417         inneriter                  += j_index_end - j_index_start;
 418
 419         /* Outer loop uses 9 flops */
 420     }
 421
 422     /* Increment number of outer iterations */
 423     outeriter        += nri;
 424
 425     /* Update outer/inner flops */
 426
 427     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*65);
 428 }
 429 /*
 430  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwLJSh_GeomP1P1_F_sse4_1_single
 431  * Electrostatics interaction: Ewald
 432  * VdW interaction:            LennardJones
 433  * Geometry:                   Particle-Particle
 434  * Calculate force/pot:        Force
 435  */
 436 void
 437 nb_kernel_ElecEwSh_VdwLJSh_GeomP1P1_F_sse4_1_single
 438                     (t_nblist * gmx_restrict                nlist,
 439                      rvec * gmx_restrict                    xx,
 440                      rvec * gmx_restrict                    ff,
 441                      t_forcerec * gmx_restrict              fr,
 442                      t_mdatoms * gmx_restrict               mdatoms,
 443                      nb_kernel_data_t * gmx_restrict        kernel_data,
 444                      t_nrnb * gmx_restrict                  nrnb)
 445 {
 446     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 447      * just 0 for non-waters.
 448      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 449      * jnr indices corresponding to data put in the four positions in the SIMD register.
 450      */
 451     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 452     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 453     int              jnrA,jnrB,jnrC,jnrD;
 454     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 455     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 456     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 457     real             rcutoff_scalar;
 458     real             *shiftvec,*fshift,*x,*f;
 459     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 460     real             scratch[4*DIM];
 461     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 462     int              vdwioffset0;
 463     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 464     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 465     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 466     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 467     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 468     real             *charge;
 469     int              nvdwtype;
 470     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 471     int              *vdwtype;
 472     real             *vdwparam;
 473     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 474     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 475     __m128i          ewitab;
 476     __m128           ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
 477     real             *ewtab;
 478     __m128           dummy_mask,cutoff_mask;
 479     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 480     __m128           one     = _mm_set1_ps(1.0);
 481     __m128           two     = _mm_set1_ps(2.0);
 482     x                = xx[0];
 483     f                = ff[0];
 484
 485     nri              = nlist->nri;
 486     iinr             = nlist->iinr;
 487     jindex           = nlist->jindex;
 488     jjnr             = nlist->jjnr;
 489     shiftidx         = nlist->shift;
 490     gid              = nlist->gid;
 491     shiftvec         = fr->shift_vec[0];
 492     fshift           = fr->fshift[0];
 493     facel            = _mm_set1_ps(fr->epsfac);
 494     charge           = mdatoms->chargeA;
 495     nvdwtype         = fr->ntype;
 496     vdwparam         = fr->nbfp;
 497     vdwtype          = mdatoms->typeA;
 498
 499     sh_ewald         = _mm_set1_ps(fr->ic->sh_ewald);
 500     ewtab            = fr->ic->tabq_coul_F;
 501     ewtabscale       = _mm_set1_ps(fr->ic->tabq_scale);
 502     ewtabhalfspace   = _mm_set1_ps(0.5/fr->ic->tabq_scale);
 503
 504     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 505     rcutoff_scalar   = fr->rcoulomb;
 506     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 507     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 508
 509     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 510     rvdw             = _mm_set1_ps(fr->rvdw);
 511
 512     /* Avoid stupid compiler warnings */
 513     jnrA = jnrB = jnrC = jnrD = 0;
 514     j_coord_offsetA = 0;
 515     j_coord_offsetB = 0;
 516     j_coord_offsetC = 0;
 517     j_coord_offsetD = 0;
 518
 519     outeriter        = 0;
 520     inneriter        = 0;
 521
 522     for(iidx=0;iidx<4*DIM;iidx++)
 523     {
 524         scratch[iidx] = 0.0;
 525     }
 526
 527     /* Start outer loop over neighborlists */
 528     for(iidx=0; iidx<nri; iidx++)
 529     {
 530         /* Load shift vector for this list */
 531         i_shift_offset   = DIM*shiftidx[iidx];
 532
 533         /* Load limits for loop over neighbors */
 534         j_index_start    = jindex[iidx];
 535         j_index_end      = jindex[iidx+1];
 536
 537         /* Get outer coordinate index */
 538         inr              = iinr[iidx];
 539         i_coord_offset   = DIM*inr;
 540
 541         /* Load i particle coords and add shift vector */
 542         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 543
 544         fix0             = _mm_setzero_ps();
 545         fiy0             = _mm_setzero_ps();
 546         fiz0             = _mm_setzero_ps();
 547
 548         /* Load parameters for i particles */
 549         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 550         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 551
 552         /* Start inner kernel loop */
 553         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 554         {
 555
 556             /* Get j neighbor index, and coordinate index */
 557             jnrA             = jjnr[jidx];
 558             jnrB             = jjnr[jidx+1];
 559             jnrC             = jjnr[jidx+2];
 560             jnrD             = jjnr[jidx+3];
 561             j_coord_offsetA  = DIM*jnrA;
 562             j_coord_offsetB  = DIM*jnrB;
 563             j_coord_offsetC  = DIM*jnrC;
 564             j_coord_offsetD  = DIM*jnrD;
 565
 566             /* load j atom coordinates */
 567             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 568                                               x+j_coord_offsetC,x+j_coord_offsetD,
 569                                               &jx0,&jy0,&jz0);
 570
 571             /* Calculate displacement vector */
 572             dx00             = _mm_sub_ps(ix0,jx0);
 573             dy00             = _mm_sub_ps(iy0,jy0);
 574             dz00             = _mm_sub_ps(iz0,jz0);
 575
 576             /* Calculate squared distance and things based on it */
 577             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 578
 579             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 580
 581             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 582
 583             /* Load parameters for j particles */
 584             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 585                                                               charge+jnrC+0,charge+jnrD+0);
 586             vdwjidx0A        = 2*vdwtype[jnrA+0];
 587             vdwjidx0B        = 2*vdwtype[jnrB+0];
 588             vdwjidx0C        = 2*vdwtype[jnrC+0];
 589             vdwjidx0D        = 2*vdwtype[jnrD+0];
 590
 591             /**************************
 592              * CALCULATE INTERACTIONS *
 593              **************************/
 594
 595             if (gmx_mm_any_lt(rsq00,rcutoff2))
 596             {
 597
 598             r00              = _mm_mul_ps(rsq00,rinv00);
 599
 600             /* Compute parameters for interactions between i and j atoms */
 601             qq00             = _mm_mul_ps(iq0,jq0);
 602             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 603                                          vdwparam+vdwioffset0+vdwjidx0B,
 604                                          vdwparam+vdwioffset0+vdwjidx0C,
 605                                          vdwparam+vdwioffset0+vdwjidx0D,
 606                                          &c6_00,&c12_00);
 607
 608             /* EWALD ELECTROSTATICS */
 609
 610             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 611             ewrt             = _mm_mul_ps(r00,ewtabscale);
 612             ewitab           = _mm_cvttps_epi32(ewrt);
 613             eweps            = _mm_sub_ps(ewrt,_mm_round_ps(ewrt, _MM_FROUND_FLOOR));
 614             gmx_mm_load_4pair_swizzle_ps(ewtab + gmx_mm_extract_epi32(ewitab,0),ewtab + gmx_mm_extract_epi32(ewitab,1),
 615                                          ewtab + gmx_mm_extract_epi32(ewitab,2),ewtab + gmx_mm_extract_epi32(ewitab,3),
 616                                          &ewtabF,&ewtabFn);
 617             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
 618             felec            = _mm_mul_ps(_mm_mul_ps(qq00,rinv00),_mm_sub_ps(rinvsq00,felec));
 619
 620             /* LENNARD-JONES DISPERSION/REPULSION */
 621
 622             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 623             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 624
 625             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 626
 627             fscal            = _mm_add_ps(felec,fvdw);
 628
 629             fscal            = _mm_and_ps(fscal,cutoff_mask);
 630
 631             /* Calculate temporary vectorial force */
 632             tx               = _mm_mul_ps(fscal,dx00);
 633             ty               = _mm_mul_ps(fscal,dy00);
 634             tz               = _mm_mul_ps(fscal,dz00);
 635
 636             /* Update vectorial force */
 637             fix0             = _mm_add_ps(fix0,tx);
 638             fiy0             = _mm_add_ps(fiy0,ty);
 639             fiz0             = _mm_add_ps(fiz0,tz);
 640
 641             fjptrA             = f+j_coord_offsetA;
 642             fjptrB             = f+j_coord_offsetB;
 643             fjptrC             = f+j_coord_offsetC;
 644             fjptrD             = f+j_coord_offsetD;
 645             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 646
 647             }
 648
 649             /* Inner loop uses 46 flops */
 650         }
 651
 652         if(jidx<j_index_end)
 653         {
 654
 655             /* Get j neighbor index, and coordinate index */
 656             jnrlistA         = jjnr[jidx];
 657             jnrlistB         = jjnr[jidx+1];
 658             jnrlistC         = jjnr[jidx+2];
 659             jnrlistD         = jjnr[jidx+3];
 660             /* Sign of each element will be negative for non-real atoms.
 661              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 662              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 663              */
 664             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 665             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 666             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 667             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 668             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 669             j_coord_offsetA  = DIM*jnrA;
 670             j_coord_offsetB  = DIM*jnrB;
 671             j_coord_offsetC  = DIM*jnrC;
 672             j_coord_offsetD  = DIM*jnrD;
 673
 674             /* load j atom coordinates */
 675             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 676                                               x+j_coord_offsetC,x+j_coord_offsetD,
 677                                               &jx0,&jy0,&jz0);
 678
 679             /* Calculate displacement vector */
 680             dx00             = _mm_sub_ps(ix0,jx0);
 681             dy00             = _mm_sub_ps(iy0,jy0);
 682             dz00             = _mm_sub_ps(iz0,jz0);
 683
 684             /* Calculate squared distance and things based on it */
 685             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 686
 687             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 688
 689             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 690
 691             /* Load parameters for j particles */
 692             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 693                                                               charge+jnrC+0,charge+jnrD+0);
 694             vdwjidx0A        = 2*vdwtype[jnrA+0];
 695             vdwjidx0B        = 2*vdwtype[jnrB+0];
 696             vdwjidx0C        = 2*vdwtype[jnrC+0];
 697             vdwjidx0D        = 2*vdwtype[jnrD+0];
 698
 699             /**************************
 700              * CALCULATE INTERACTIONS *
 701              **************************/
 702
 703             if (gmx_mm_any_lt(rsq00,rcutoff2))
 704             {
 705
 706             r00              = _mm_mul_ps(rsq00,rinv00);
 707             r00              = _mm_andnot_ps(dummy_mask,r00);
 708
 709             /* Compute parameters for interactions between i and j atoms */
 710             qq00             = _mm_mul_ps(iq0,jq0);
 711             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 712                                          vdwparam+vdwioffset0+vdwjidx0B,
 713                                          vdwparam+vdwioffset0+vdwjidx0C,
 714                                          vdwparam+vdwioffset0+vdwjidx0D,
 715                                          &c6_00,&c12_00);
 716
 717             /* EWALD ELECTROSTATICS */
 718
 719             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 720             ewrt             = _mm_mul_ps(r00,ewtabscale);
 721             ewitab           = _mm_cvttps_epi32(ewrt);
 722             eweps            = _mm_sub_ps(ewrt,_mm_round_ps(ewrt, _MM_FROUND_FLOOR));
 723             gmx_mm_load_4pair_swizzle_ps(ewtab + gmx_mm_extract_epi32(ewitab,0),ewtab + gmx_mm_extract_epi32(ewitab,1),
 724                                          ewtab + gmx_mm_extract_epi32(ewitab,2),ewtab + gmx_mm_extract_epi32(ewitab,3),
 725                                          &ewtabF,&ewtabFn);
 726             felec            = _mm_add_ps(_mm_mul_ps( _mm_sub_ps(one,eweps),ewtabF),_mm_mul_ps(eweps,ewtabFn));
 727             felec            = _mm_mul_ps(_mm_mul_ps(qq00,rinv00),_mm_sub_ps(rinvsq00,felec));
 728
 729             /* LENNARD-JONES DISPERSION/REPULSION */
 730
 731             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 732             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
 733
 734             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 735
 736             fscal            = _mm_add_ps(felec,fvdw);
 737
 738             fscal            = _mm_and_ps(fscal,cutoff_mask);
 739
 740             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 741
 742             /* Calculate temporary vectorial force */
 743             tx               = _mm_mul_ps(fscal,dx00);
 744             ty               = _mm_mul_ps(fscal,dy00);
 745             tz               = _mm_mul_ps(fscal,dz00);
 746
 747             /* Update vectorial force */
 748             fix0             = _mm_add_ps(fix0,tx);
 749             fiy0             = _mm_add_ps(fiy0,ty);
 750             fiz0             = _mm_add_ps(fiz0,tz);
 751
 752             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 753             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 754             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 755             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 756             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,tx,ty,tz);
 757
 758             }
 759
 760             /* Inner loop uses 47 flops */
 761         }
 762
 763         /* End of innermost loop */
 764
 765         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 766                                               f+i_coord_offset,fshift+i_shift_offset);
 767
 768         /* Increment number of inner iterations */
 769         inneriter                  += j_index_end - j_index_start;
 770
 771         /* Outer loop uses 7 flops */
 772     }
 773
 774     /* Increment number of outer iterations */
 775     outeriter        += nri;
 776
 777     /* Update outer/inner flops */
 778
 779     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*47);
 780 }