src/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_avx_128_fma_single.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_128_fma_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_avx_128_fma_single.h"
  34 #include "kernelutil_x86_avx_128_fma_single.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_VF_avx_128_fma_single
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            LennardJones
  40  * Geometry:                   Particle-Particle
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_VF_avx_128_fma_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 AVX_128, 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           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     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
  83     real             rswitch_scalar,d_scalar;
  84     __m128           dummy_mask,cutoff_mask;
  85     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
  86     __m128           one     = _mm_set1_ps(1.0);
  87     __m128           two     = _mm_set1_ps(2.0);
  88     x                = xx[0];
  89     f                = ff[0];
  90
  91     nri              = nlist->nri;
  92     iinr             = nlist->iinr;
  93     jindex           = nlist->jindex;
  94     jjnr             = nlist->jjnr;
  95     shiftidx         = nlist->shift;
  96     gid              = nlist->gid;
  97     shiftvec         = fr->shift_vec[0];
  98     fshift           = fr->fshift[0];
  99     facel            = _mm_set1_ps(fr->epsfac);
 100     charge           = mdatoms->chargeA;
 101     krf              = _mm_set1_ps(fr->ic->k_rf);
 102     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 103     crf              = _mm_set1_ps(fr->ic->c_rf);
 104     nvdwtype         = fr->ntype;
 105     vdwparam         = fr->nbfp;
 106     vdwtype          = mdatoms->typeA;
 107
 108     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 109     rcutoff_scalar   = fr->rcoulomb;
 110     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 111     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 112
 113     rswitch_scalar   = fr->rvdw_switch;
 114     rswitch          = _mm_set1_ps(rswitch_scalar);
 115     /* Setup switch parameters */
 116     d_scalar         = rcutoff_scalar-rswitch_scalar;
 117     d                = _mm_set1_ps(d_scalar);
 118     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 119     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 120     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 121     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 122     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 123     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 124
 125     /* Avoid stupid compiler warnings */
 126     jnrA = jnrB = jnrC = jnrD = 0;
 127     j_coord_offsetA = 0;
 128     j_coord_offsetB = 0;
 129     j_coord_offsetC = 0;
 130     j_coord_offsetD = 0;
 131
 132     outeriter        = 0;
 133     inneriter        = 0;
 134
 135     for(iidx=0;iidx<4*DIM;iidx++)
 136     {
 137         scratch[iidx] = 0.0;
 138     }
 139
 140     /* Start outer loop over neighborlists */
 141     for(iidx=0; iidx<nri; iidx++)
 142     {
 143         /* Load shift vector for this list */
 144         i_shift_offset   = DIM*shiftidx[iidx];
 145
 146         /* Load limits for loop over neighbors */
 147         j_index_start    = jindex[iidx];
 148         j_index_end      = jindex[iidx+1];
 149
 150         /* Get outer coordinate index */
 151         inr              = iinr[iidx];
 152         i_coord_offset   = DIM*inr;
 153
 154         /* Load i particle coords and add shift vector */
 155         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 156
 157         fix0             = _mm_setzero_ps();
 158         fiy0             = _mm_setzero_ps();
 159         fiz0             = _mm_setzero_ps();
 160
 161         /* Load parameters for i particles */
 162         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 163         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 164
 165         /* Reset potential sums */
 166         velecsum         = _mm_setzero_ps();
 167         vvdwsum          = _mm_setzero_ps();
 168
 169         /* Start inner kernel loop */
 170         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 171         {
 172
 173             /* Get j neighbor index, and coordinate index */
 174             jnrA             = jjnr[jidx];
 175             jnrB             = jjnr[jidx+1];
 176             jnrC             = jjnr[jidx+2];
 177             jnrD             = jjnr[jidx+3];
 178             j_coord_offsetA  = DIM*jnrA;
 179             j_coord_offsetB  = DIM*jnrB;
 180             j_coord_offsetC  = DIM*jnrC;
 181             j_coord_offsetD  = DIM*jnrD;
 182
 183             /* load j atom coordinates */
 184             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 185                                               x+j_coord_offsetC,x+j_coord_offsetD,
 186                                               &jx0,&jy0,&jz0);
 187
 188             /* Calculate displacement vector */
 189             dx00             = _mm_sub_ps(ix0,jx0);
 190             dy00             = _mm_sub_ps(iy0,jy0);
 191             dz00             = _mm_sub_ps(iz0,jz0);
 192
 193             /* Calculate squared distance and things based on it */
 194             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 195
 196             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 197
 198             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 199
 200             /* Load parameters for j particles */
 201             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 202                                                               charge+jnrC+0,charge+jnrD+0);
 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             qq00             = _mm_mul_ps(iq0,jq0);
 219             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 220                                          vdwparam+vdwioffset0+vdwjidx0B,
 221                                          vdwparam+vdwioffset0+vdwjidx0C,
 222                                          vdwparam+vdwioffset0+vdwjidx0D,
 223                                          &c6_00,&c12_00);
 224
 225             /* REACTION-FIELD ELECTROSTATICS */
 226             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 227             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 228
 229             /* LENNARD-JONES DISPERSION/REPULSION */
 230
 231             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 232             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 233             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 234             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 235             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 236
 237             d                = _mm_sub_ps(r00,rswitch);
 238             d                = _mm_max_ps(d,_mm_setzero_ps());
 239             d2               = _mm_mul_ps(d,d);
 240             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 241
 242             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 243
 244             /* Evaluate switch function */
 245             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 246             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 247             vvdw             = _mm_mul_ps(vvdw,sw);
 248             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 249
 250             /* Update potential sum for this i atom from the interaction with this j atom. */
 251             velec            = _mm_and_ps(velec,cutoff_mask);
 252             velecsum         = _mm_add_ps(velecsum,velec);
 253             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 254             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 255
 256             fscal            = _mm_add_ps(felec,fvdw);
 257
 258             fscal            = _mm_and_ps(fscal,cutoff_mask);
 259
 260              /* Update vectorial force */
 261             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 262             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 263             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 264
 265             fjptrA             = f+j_coord_offsetA;
 266             fjptrB             = f+j_coord_offsetB;
 267             fjptrC             = f+j_coord_offsetC;
 268             fjptrD             = f+j_coord_offsetD;
 269             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 270                                                    _mm_mul_ps(dx00,fscal),
 271                                                    _mm_mul_ps(dy00,fscal),
 272                                                    _mm_mul_ps(dz00,fscal));
 273
 274             }
 275
 276             /* Inner loop uses 73 flops */
 277         }
 278
 279         if(jidx<j_index_end)
 280         {
 281
 282             /* Get j neighbor index, and coordinate index */
 283             jnrlistA         = jjnr[jidx];
 284             jnrlistB         = jjnr[jidx+1];
 285             jnrlistC         = jjnr[jidx+2];
 286             jnrlistD         = jjnr[jidx+3];
 287             /* Sign of each element will be negative for non-real atoms.
 288              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 289              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 290              */
 291             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 292             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 293             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 294             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 295             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 296             j_coord_offsetA  = DIM*jnrA;
 297             j_coord_offsetB  = DIM*jnrB;
 298             j_coord_offsetC  = DIM*jnrC;
 299             j_coord_offsetD  = DIM*jnrD;
 300
 301             /* load j atom coordinates */
 302             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 303                                               x+j_coord_offsetC,x+j_coord_offsetD,
 304                                               &jx0,&jy0,&jz0);
 305
 306             /* Calculate displacement vector */
 307             dx00             = _mm_sub_ps(ix0,jx0);
 308             dy00             = _mm_sub_ps(iy0,jy0);
 309             dz00             = _mm_sub_ps(iz0,jz0);
 310
 311             /* Calculate squared distance and things based on it */
 312             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 313
 314             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 315
 316             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 317
 318             /* Load parameters for j particles */
 319             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 320                                                               charge+jnrC+0,charge+jnrD+0);
 321             vdwjidx0A        = 2*vdwtype[jnrA+0];
 322             vdwjidx0B        = 2*vdwtype[jnrB+0];
 323             vdwjidx0C        = 2*vdwtype[jnrC+0];
 324             vdwjidx0D        = 2*vdwtype[jnrD+0];
 325
 326             /**************************
 327              * CALCULATE INTERACTIONS *
 328              **************************/
 329
 330             if (gmx_mm_any_lt(rsq00,rcutoff2))
 331             {
 332
 333             r00              = _mm_mul_ps(rsq00,rinv00);
 334             r00              = _mm_andnot_ps(dummy_mask,r00);
 335
 336             /* Compute parameters for interactions between i and j atoms */
 337             qq00             = _mm_mul_ps(iq0,jq0);
 338             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 339                                          vdwparam+vdwioffset0+vdwjidx0B,
 340                                          vdwparam+vdwioffset0+vdwjidx0C,
 341                                          vdwparam+vdwioffset0+vdwjidx0D,
 342                                          &c6_00,&c12_00);
 343
 344             /* REACTION-FIELD ELECTROSTATICS */
 345             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 346             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 347
 348             /* LENNARD-JONES DISPERSION/REPULSION */
 349
 350             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 351             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 352             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 353             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 354             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 355
 356             d                = _mm_sub_ps(r00,rswitch);
 357             d                = _mm_max_ps(d,_mm_setzero_ps());
 358             d2               = _mm_mul_ps(d,d);
 359             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 360
 361             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 362
 363             /* Evaluate switch function */
 364             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 365             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 366             vvdw             = _mm_mul_ps(vvdw,sw);
 367             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 368
 369             /* Update potential sum for this i atom from the interaction with this j atom. */
 370             velec            = _mm_and_ps(velec,cutoff_mask);
 371             velec            = _mm_andnot_ps(dummy_mask,velec);
 372             velecsum         = _mm_add_ps(velecsum,velec);
 373             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 374             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 375             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 376
 377             fscal            = _mm_add_ps(felec,fvdw);
 378
 379             fscal            = _mm_and_ps(fscal,cutoff_mask);
 380
 381             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 382
 383              /* Update vectorial force */
 384             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 385             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 386             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 387
 388             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 389             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 390             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 391             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 392             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 393                                                    _mm_mul_ps(dx00,fscal),
 394                                                    _mm_mul_ps(dy00,fscal),
 395                                                    _mm_mul_ps(dz00,fscal));
 396
 397             }
 398
 399             /* Inner loop uses 74 flops */
 400         }
 401
 402         /* End of innermost loop */
 403
 404         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 405                                               f+i_coord_offset,fshift+i_shift_offset);
 406
 407         ggid                        = gid[iidx];
 408         /* Update potential energies */
 409         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 410         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 411
 412         /* Increment number of inner iterations */
 413         inneriter                  += j_index_end - j_index_start;
 414
 415         /* Outer loop uses 9 flops */
 416     }
 417
 418     /* Increment number of outer iterations */
 419     outeriter        += nri;
 420
 421     /* Update outer/inner flops */
 422
 423     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_VF,outeriter*9 + inneriter*74);
 424 }
 425 /*
 426  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_F_avx_128_fma_single
 427  * Electrostatics interaction: ReactionField
 428  * VdW interaction:            LennardJones
 429  * Geometry:                   Particle-Particle
 430  * Calculate force/pot:        Force
 431  */
 432 void
 433 nb_kernel_ElecRFCut_VdwLJSw_GeomP1P1_F_avx_128_fma_single
 434                     (t_nblist * gmx_restrict                nlist,
 435                      rvec * gmx_restrict                    xx,
 436                      rvec * gmx_restrict                    ff,
 437                      t_forcerec * gmx_restrict              fr,
 438                      t_mdatoms * gmx_restrict               mdatoms,
 439                      nb_kernel_data_t * gmx_restrict        kernel_data,
 440                      t_nrnb * gmx_restrict                  nrnb)
 441 {
 442     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 443      * just 0 for non-waters.
 444      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 445      * jnr indices corresponding to data put in the four positions in the SIMD register.
 446      */
 447     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 448     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 449     int              jnrA,jnrB,jnrC,jnrD;
 450     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 451     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 452     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 453     real             rcutoff_scalar;
 454     real             *shiftvec,*fshift,*x,*f;
 455     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 456     real             scratch[4*DIM];
 457     __m128           fscal,rcutoff,rcutoff2,jidxall;
 458     int              vdwioffset0;
 459     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 460     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 461     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 462     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 463     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 464     real             *charge;
 465     int              nvdwtype;
 466     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 467     int              *vdwtype;
 468     real             *vdwparam;
 469     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 470     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 471     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 472     real             rswitch_scalar,d_scalar;
 473     __m128           dummy_mask,cutoff_mask;
 474     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 475     __m128           one     = _mm_set1_ps(1.0);
 476     __m128           two     = _mm_set1_ps(2.0);
 477     x                = xx[0];
 478     f                = ff[0];
 479
 480     nri              = nlist->nri;
 481     iinr             = nlist->iinr;
 482     jindex           = nlist->jindex;
 483     jjnr             = nlist->jjnr;
 484     shiftidx         = nlist->shift;
 485     gid              = nlist->gid;
 486     shiftvec         = fr->shift_vec[0];
 487     fshift           = fr->fshift[0];
 488     facel            = _mm_set1_ps(fr->epsfac);
 489     charge           = mdatoms->chargeA;
 490     krf              = _mm_set1_ps(fr->ic->k_rf);
 491     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 492     crf              = _mm_set1_ps(fr->ic->c_rf);
 493     nvdwtype         = fr->ntype;
 494     vdwparam         = fr->nbfp;
 495     vdwtype          = mdatoms->typeA;
 496
 497     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 498     rcutoff_scalar   = fr->rcoulomb;
 499     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 500     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 501
 502     rswitch_scalar   = fr->rvdw_switch;
 503     rswitch          = _mm_set1_ps(rswitch_scalar);
 504     /* Setup switch parameters */
 505     d_scalar         = rcutoff_scalar-rswitch_scalar;
 506     d                = _mm_set1_ps(d_scalar);
 507     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 508     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 509     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 510     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 511     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 512     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 513
 514     /* Avoid stupid compiler warnings */
 515     jnrA = jnrB = jnrC = jnrD = 0;
 516     j_coord_offsetA = 0;
 517     j_coord_offsetB = 0;
 518     j_coord_offsetC = 0;
 519     j_coord_offsetD = 0;
 520
 521     outeriter        = 0;
 522     inneriter        = 0;
 523
 524     for(iidx=0;iidx<4*DIM;iidx++)
 525     {
 526         scratch[iidx] = 0.0;
 527     }
 528
 529     /* Start outer loop over neighborlists */
 530     for(iidx=0; iidx<nri; iidx++)
 531     {
 532         /* Load shift vector for this list */
 533         i_shift_offset   = DIM*shiftidx[iidx];
 534
 535         /* Load limits for loop over neighbors */
 536         j_index_start    = jindex[iidx];
 537         j_index_end      = jindex[iidx+1];
 538
 539         /* Get outer coordinate index */
 540         inr              = iinr[iidx];
 541         i_coord_offset   = DIM*inr;
 542
 543         /* Load i particle coords and add shift vector */
 544         gmx_mm_load_shift_and_1rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,&ix0,&iy0,&iz0);
 545
 546         fix0             = _mm_setzero_ps();
 547         fiy0             = _mm_setzero_ps();
 548         fiz0             = _mm_setzero_ps();
 549
 550         /* Load parameters for i particles */
 551         iq0              = _mm_mul_ps(facel,_mm_load1_ps(charge+inr+0));
 552         vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 553
 554         /* Start inner kernel loop */
 555         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 556         {
 557
 558             /* Get j neighbor index, and coordinate index */
 559             jnrA             = jjnr[jidx];
 560             jnrB             = jjnr[jidx+1];
 561             jnrC             = jjnr[jidx+2];
 562             jnrD             = jjnr[jidx+3];
 563             j_coord_offsetA  = DIM*jnrA;
 564             j_coord_offsetB  = DIM*jnrB;
 565             j_coord_offsetC  = DIM*jnrC;
 566             j_coord_offsetD  = DIM*jnrD;
 567
 568             /* load j atom coordinates */
 569             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 570                                               x+j_coord_offsetC,x+j_coord_offsetD,
 571                                               &jx0,&jy0,&jz0);
 572
 573             /* Calculate displacement vector */
 574             dx00             = _mm_sub_ps(ix0,jx0);
 575             dy00             = _mm_sub_ps(iy0,jy0);
 576             dz00             = _mm_sub_ps(iz0,jz0);
 577
 578             /* Calculate squared distance and things based on it */
 579             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 580
 581             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 582
 583             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 584
 585             /* Load parameters for j particles */
 586             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 587                                                               charge+jnrC+0,charge+jnrD+0);
 588             vdwjidx0A        = 2*vdwtype[jnrA+0];
 589             vdwjidx0B        = 2*vdwtype[jnrB+0];
 590             vdwjidx0C        = 2*vdwtype[jnrC+0];
 591             vdwjidx0D        = 2*vdwtype[jnrD+0];
 592
 593             /**************************
 594              * CALCULATE INTERACTIONS *
 595              **************************/
 596
 597             if (gmx_mm_any_lt(rsq00,rcutoff2))
 598             {
 599
 600             r00              = _mm_mul_ps(rsq00,rinv00);
 601
 602             /* Compute parameters for interactions between i and j atoms */
 603             qq00             = _mm_mul_ps(iq0,jq0);
 604             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 605                                          vdwparam+vdwioffset0+vdwjidx0B,
 606                                          vdwparam+vdwioffset0+vdwjidx0C,
 607                                          vdwparam+vdwioffset0+vdwjidx0D,
 608                                          &c6_00,&c12_00);
 609
 610             /* REACTION-FIELD ELECTROSTATICS */
 611             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 612
 613             /* LENNARD-JONES DISPERSION/REPULSION */
 614
 615             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 616             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 617             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 618             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 619             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 620
 621             d                = _mm_sub_ps(r00,rswitch);
 622             d                = _mm_max_ps(d,_mm_setzero_ps());
 623             d2               = _mm_mul_ps(d,d);
 624             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 625
 626             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 627
 628             /* Evaluate switch function */
 629             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 630             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 631             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 632
 633             fscal            = _mm_add_ps(felec,fvdw);
 634
 635             fscal            = _mm_and_ps(fscal,cutoff_mask);
 636
 637              /* Update vectorial force */
 638             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 639             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 640             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 641
 642             fjptrA             = f+j_coord_offsetA;
 643             fjptrB             = f+j_coord_offsetB;
 644             fjptrC             = f+j_coord_offsetC;
 645             fjptrD             = f+j_coord_offsetD;
 646             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 647                                                    _mm_mul_ps(dx00,fscal),
 648                                                    _mm_mul_ps(dy00,fscal),
 649                                                    _mm_mul_ps(dz00,fscal));
 650
 651             }
 652
 653             /* Inner loop uses 64 flops */
 654         }
 655
 656         if(jidx<j_index_end)
 657         {
 658
 659             /* Get j neighbor index, and coordinate index */
 660             jnrlistA         = jjnr[jidx];
 661             jnrlistB         = jjnr[jidx+1];
 662             jnrlistC         = jjnr[jidx+2];
 663             jnrlistD         = jjnr[jidx+3];
 664             /* Sign of each element will be negative for non-real atoms.
 665              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 666              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 667              */
 668             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 669             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 670             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 671             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 672             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 673             j_coord_offsetA  = DIM*jnrA;
 674             j_coord_offsetB  = DIM*jnrB;
 675             j_coord_offsetC  = DIM*jnrC;
 676             j_coord_offsetD  = DIM*jnrD;
 677
 678             /* load j atom coordinates */
 679             gmx_mm_load_1rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 680                                               x+j_coord_offsetC,x+j_coord_offsetD,
 681                                               &jx0,&jy0,&jz0);
 682
 683             /* Calculate displacement vector */
 684             dx00             = _mm_sub_ps(ix0,jx0);
 685             dy00             = _mm_sub_ps(iy0,jy0);
 686             dz00             = _mm_sub_ps(iz0,jz0);
 687
 688             /* Calculate squared distance and things based on it */
 689             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 690
 691             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 692
 693             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 694
 695             /* Load parameters for j particles */
 696             jq0              = gmx_mm_load_4real_swizzle_ps(charge+jnrA+0,charge+jnrB+0,
 697                                                               charge+jnrC+0,charge+jnrD+0);
 698             vdwjidx0A        = 2*vdwtype[jnrA+0];
 699             vdwjidx0B        = 2*vdwtype[jnrB+0];
 700             vdwjidx0C        = 2*vdwtype[jnrC+0];
 701             vdwjidx0D        = 2*vdwtype[jnrD+0];
 702
 703             /**************************
 704              * CALCULATE INTERACTIONS *
 705              **************************/
 706
 707             if (gmx_mm_any_lt(rsq00,rcutoff2))
 708             {
 709
 710             r00              = _mm_mul_ps(rsq00,rinv00);
 711             r00              = _mm_andnot_ps(dummy_mask,r00);
 712
 713             /* Compute parameters for interactions between i and j atoms */
 714             qq00             = _mm_mul_ps(iq0,jq0);
 715             gmx_mm_load_4pair_swizzle_ps(vdwparam+vdwioffset0+vdwjidx0A,
 716                                          vdwparam+vdwioffset0+vdwjidx0B,
 717                                          vdwparam+vdwioffset0+vdwjidx0C,
 718                                          vdwparam+vdwioffset0+vdwjidx0D,
 719                                          &c6_00,&c12_00);
 720
 721             /* REACTION-FIELD ELECTROSTATICS */
 722             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 723
 724             /* LENNARD-JONES DISPERSION/REPULSION */
 725
 726             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 727             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 728             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 729             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 730             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 731
 732             d                = _mm_sub_ps(r00,rswitch);
 733             d                = _mm_max_ps(d,_mm_setzero_ps());
 734             d2               = _mm_mul_ps(d,d);
 735             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 736
 737             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 738
 739             /* Evaluate switch function */
 740             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 741             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 742             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 743
 744             fscal            = _mm_add_ps(felec,fvdw);
 745
 746             fscal            = _mm_and_ps(fscal,cutoff_mask);
 747
 748             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 749
 750              /* Update vectorial force */
 751             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 752             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 753             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 754
 755             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 756             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 757             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 758             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 759             gmx_mm_decrement_1rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 760                                                    _mm_mul_ps(dx00,fscal),
 761                                                    _mm_mul_ps(dy00,fscal),
 762                                                    _mm_mul_ps(dz00,fscal));
 763
 764             }
 765
 766             /* Inner loop uses 65 flops */
 767         }
 768
 769         /* End of innermost loop */
 770
 771         gmx_mm_update_iforce_1atom_swizzle_ps(fix0,fiy0,fiz0,
 772                                               f+i_coord_offset,fshift+i_shift_offset);
 773
 774         /* Increment number of inner iterations */
 775         inneriter                  += j_index_end - j_index_start;
 776
 777         /* Outer loop uses 7 flops */
 778     }
 779
 780     /* Increment number of outer iterations */
 781     outeriter        += nri;
 782
 783     /* Update outer/inner flops */
 784
 785     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_F,outeriter*7 + inneriter*65);
 786 }