src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_ElecCoul_VdwLJ_GeomW4W4_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_ElecCoul_VdwLJ_GeomW4W4_VF_sse4_1_single
  52  * Electrostatics interaction: Coulomb
  53  * VdW interaction:            LennardJones
  54  * Geometry:                   Water4-Water4
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecCoul_VdwLJ_GeomW4W4_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              vdwioffset1;
  86     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  87     int              vdwioffset2;
  88     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  89     int              vdwioffset3;
  90     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  91     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  92     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  93     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  94     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  95     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  96     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  97     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  98     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  99     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 100     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 101     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 102     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 103     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 104     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 105     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 106     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 107     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 108     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 109     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 110     real             *charge;
 111     int              nvdwtype;
 112     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 113     int              *vdwtype;
 114     real             *vdwparam;
 115     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 116     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 117     __m128           dummy_mask,cutoff_mask;
 118     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 119     __m128           one     = _mm_set1_ps(1.0);
 120     __m128           two     = _mm_set1_ps(2.0);
 121     x                = xx[0];
 122     f                = ff[0];
 123
 124     nri              = nlist->nri;
 125     iinr             = nlist->iinr;
 126     jindex           = nlist->jindex;
 127     jjnr             = nlist->jjnr;
 128     shiftidx         = nlist->shift;
 129     gid              = nlist->gid;
 130     shiftvec         = fr->shift_vec[0];
 131     fshift           = fr->fshift[0];
 132     facel            = _mm_set1_ps(fr->epsfac);
 133     charge           = mdatoms->chargeA;
 134     nvdwtype         = fr->ntype;
 135     vdwparam         = fr->nbfp;
 136     vdwtype          = mdatoms->typeA;
 137
 138     /* Setup water-specific parameters */
 139     inr              = nlist->iinr[0];
 140     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 141     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 142     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 143     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 144
 145     jq1              = _mm_set1_ps(charge[inr+1]);
 146     jq2              = _mm_set1_ps(charge[inr+2]);
 147     jq3              = _mm_set1_ps(charge[inr+3]);
 148     vdwjidx0A        = 2*vdwtype[inr+0];
 149     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 150     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 151     qq11             = _mm_mul_ps(iq1,jq1);
 152     qq12             = _mm_mul_ps(iq1,jq2);
 153     qq13             = _mm_mul_ps(iq1,jq3);
 154     qq21             = _mm_mul_ps(iq2,jq1);
 155     qq22             = _mm_mul_ps(iq2,jq2);
 156     qq23             = _mm_mul_ps(iq2,jq3);
 157     qq31             = _mm_mul_ps(iq3,jq1);
 158     qq32             = _mm_mul_ps(iq3,jq2);
 159     qq33             = _mm_mul_ps(iq3,jq3);
 160
 161     /* Avoid stupid compiler warnings */
 162     jnrA = jnrB = jnrC = jnrD = 0;
 163     j_coord_offsetA = 0;
 164     j_coord_offsetB = 0;
 165     j_coord_offsetC = 0;
 166     j_coord_offsetD = 0;
 167
 168     outeriter        = 0;
 169     inneriter        = 0;
 170
 171     for(iidx=0;iidx<4*DIM;iidx++)
 172     {
 173         scratch[iidx] = 0.0;
 174     }
 175
 176     /* Start outer loop over neighborlists */
 177     for(iidx=0; iidx<nri; iidx++)
 178     {
 179         /* Load shift vector for this list */
 180         i_shift_offset   = DIM*shiftidx[iidx];
 181
 182         /* Load limits for loop over neighbors */
 183         j_index_start    = jindex[iidx];
 184         j_index_end      = jindex[iidx+1];
 185
 186         /* Get outer coordinate index */
 187         inr              = iinr[iidx];
 188         i_coord_offset   = DIM*inr;
 189
 190         /* Load i particle coords and add shift vector */
 191         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 192                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 193
 194         fix0             = _mm_setzero_ps();
 195         fiy0             = _mm_setzero_ps();
 196         fiz0             = _mm_setzero_ps();
 197         fix1             = _mm_setzero_ps();
 198         fiy1             = _mm_setzero_ps();
 199         fiz1             = _mm_setzero_ps();
 200         fix2             = _mm_setzero_ps();
 201         fiy2             = _mm_setzero_ps();
 202         fiz2             = _mm_setzero_ps();
 203         fix3             = _mm_setzero_ps();
 204         fiy3             = _mm_setzero_ps();
 205         fiz3             = _mm_setzero_ps();
 206
 207         /* Reset potential sums */
 208         velecsum         = _mm_setzero_ps();
 209         vvdwsum          = _mm_setzero_ps();
 210
 211         /* Start inner kernel loop */
 212         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 213         {
 214
 215             /* Get j neighbor index, and coordinate index */
 216             jnrA             = jjnr[jidx];
 217             jnrB             = jjnr[jidx+1];
 218             jnrC             = jjnr[jidx+2];
 219             jnrD             = jjnr[jidx+3];
 220             j_coord_offsetA  = DIM*jnrA;
 221             j_coord_offsetB  = DIM*jnrB;
 222             j_coord_offsetC  = DIM*jnrC;
 223             j_coord_offsetD  = DIM*jnrD;
 224
 225             /* load j atom coordinates */
 226             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 227                                               x+j_coord_offsetC,x+j_coord_offsetD,
 228                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 229                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 230
 231             /* Calculate displacement vector */
 232             dx00             = _mm_sub_ps(ix0,jx0);
 233             dy00             = _mm_sub_ps(iy0,jy0);
 234             dz00             = _mm_sub_ps(iz0,jz0);
 235             dx11             = _mm_sub_ps(ix1,jx1);
 236             dy11             = _mm_sub_ps(iy1,jy1);
 237             dz11             = _mm_sub_ps(iz1,jz1);
 238             dx12             = _mm_sub_ps(ix1,jx2);
 239             dy12             = _mm_sub_ps(iy1,jy2);
 240             dz12             = _mm_sub_ps(iz1,jz2);
 241             dx13             = _mm_sub_ps(ix1,jx3);
 242             dy13             = _mm_sub_ps(iy1,jy3);
 243             dz13             = _mm_sub_ps(iz1,jz3);
 244             dx21             = _mm_sub_ps(ix2,jx1);
 245             dy21             = _mm_sub_ps(iy2,jy1);
 246             dz21             = _mm_sub_ps(iz2,jz1);
 247             dx22             = _mm_sub_ps(ix2,jx2);
 248             dy22             = _mm_sub_ps(iy2,jy2);
 249             dz22             = _mm_sub_ps(iz2,jz2);
 250             dx23             = _mm_sub_ps(ix2,jx3);
 251             dy23             = _mm_sub_ps(iy2,jy3);
 252             dz23             = _mm_sub_ps(iz2,jz3);
 253             dx31             = _mm_sub_ps(ix3,jx1);
 254             dy31             = _mm_sub_ps(iy3,jy1);
 255             dz31             = _mm_sub_ps(iz3,jz1);
 256             dx32             = _mm_sub_ps(ix3,jx2);
 257             dy32             = _mm_sub_ps(iy3,jy2);
 258             dz32             = _mm_sub_ps(iz3,jz2);
 259             dx33             = _mm_sub_ps(ix3,jx3);
 260             dy33             = _mm_sub_ps(iy3,jy3);
 261             dz33             = _mm_sub_ps(iz3,jz3);
 262
 263             /* Calculate squared distance and things based on it */
 264             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 265             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 266             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 267             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 268             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 269             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 270             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 271             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 272             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 273             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 274
 275             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 276             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 277             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 278             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 279             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 280             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 281             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 282             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 283             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 284
 285             rinvsq00         = gmx_mm_inv_ps(rsq00);
 286             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 287             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 288             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 289             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 290             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 291             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 292             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 293             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 294             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 295
 296             fjx0             = _mm_setzero_ps();
 297             fjy0             = _mm_setzero_ps();
 298             fjz0             = _mm_setzero_ps();
 299             fjx1             = _mm_setzero_ps();
 300             fjy1             = _mm_setzero_ps();
 301             fjz1             = _mm_setzero_ps();
 302             fjx2             = _mm_setzero_ps();
 303             fjy2             = _mm_setzero_ps();
 304             fjz2             = _mm_setzero_ps();
 305             fjx3             = _mm_setzero_ps();
 306             fjy3             = _mm_setzero_ps();
 307             fjz3             = _mm_setzero_ps();
 308
 309             /**************************
 310              * CALCULATE INTERACTIONS *
 311              **************************/
 312
 313             /* LENNARD-JONES DISPERSION/REPULSION */
 314
 315             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 316             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 317             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 318             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 319             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 320
 321             /* Update potential sum for this i atom from the interaction with this j atom. */
 322             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 323
 324             fscal            = fvdw;
 325
 326             /* Calculate temporary vectorial force */
 327             tx               = _mm_mul_ps(fscal,dx00);
 328             ty               = _mm_mul_ps(fscal,dy00);
 329             tz               = _mm_mul_ps(fscal,dz00);
 330
 331             /* Update vectorial force */
 332             fix0             = _mm_add_ps(fix0,tx);
 333             fiy0             = _mm_add_ps(fiy0,ty);
 334             fiz0             = _mm_add_ps(fiz0,tz);
 335
 336             fjx0             = _mm_add_ps(fjx0,tx);
 337             fjy0             = _mm_add_ps(fjy0,ty);
 338             fjz0             = _mm_add_ps(fjz0,tz);
 339
 340             /**************************
 341              * CALCULATE INTERACTIONS *
 342              **************************/
 343
 344             /* COULOMB ELECTROSTATICS */
 345             velec            = _mm_mul_ps(qq11,rinv11);
 346             felec            = _mm_mul_ps(velec,rinvsq11);
 347
 348             /* Update potential sum for this i atom from the interaction with this j atom. */
 349             velecsum         = _mm_add_ps(velecsum,velec);
 350
 351             fscal            = felec;
 352
 353             /* Calculate temporary vectorial force */
 354             tx               = _mm_mul_ps(fscal,dx11);
 355             ty               = _mm_mul_ps(fscal,dy11);
 356             tz               = _mm_mul_ps(fscal,dz11);
 357
 358             /* Update vectorial force */
 359             fix1             = _mm_add_ps(fix1,tx);
 360             fiy1             = _mm_add_ps(fiy1,ty);
 361             fiz1             = _mm_add_ps(fiz1,tz);
 362
 363             fjx1             = _mm_add_ps(fjx1,tx);
 364             fjy1             = _mm_add_ps(fjy1,ty);
 365             fjz1             = _mm_add_ps(fjz1,tz);
 366
 367             /**************************
 368              * CALCULATE INTERACTIONS *
 369              **************************/
 370
 371             /* COULOMB ELECTROSTATICS */
 372             velec            = _mm_mul_ps(qq12,rinv12);
 373             felec            = _mm_mul_ps(velec,rinvsq12);
 374
 375             /* Update potential sum for this i atom from the interaction with this j atom. */
 376             velecsum         = _mm_add_ps(velecsum,velec);
 377
 378             fscal            = felec;
 379
 380             /* Calculate temporary vectorial force */
 381             tx               = _mm_mul_ps(fscal,dx12);
 382             ty               = _mm_mul_ps(fscal,dy12);
 383             tz               = _mm_mul_ps(fscal,dz12);
 384
 385             /* Update vectorial force */
 386             fix1             = _mm_add_ps(fix1,tx);
 387             fiy1             = _mm_add_ps(fiy1,ty);
 388             fiz1             = _mm_add_ps(fiz1,tz);
 389
 390             fjx2             = _mm_add_ps(fjx2,tx);
 391             fjy2             = _mm_add_ps(fjy2,ty);
 392             fjz2             = _mm_add_ps(fjz2,tz);
 393
 394             /**************************
 395              * CALCULATE INTERACTIONS *
 396              **************************/
 397
 398             /* COULOMB ELECTROSTATICS */
 399             velec            = _mm_mul_ps(qq13,rinv13);
 400             felec            = _mm_mul_ps(velec,rinvsq13);
 401
 402             /* Update potential sum for this i atom from the interaction with this j atom. */
 403             velecsum         = _mm_add_ps(velecsum,velec);
 404
 405             fscal            = felec;
 406
 407             /* Calculate temporary vectorial force */
 408             tx               = _mm_mul_ps(fscal,dx13);
 409             ty               = _mm_mul_ps(fscal,dy13);
 410             tz               = _mm_mul_ps(fscal,dz13);
 411
 412             /* Update vectorial force */
 413             fix1             = _mm_add_ps(fix1,tx);
 414             fiy1             = _mm_add_ps(fiy1,ty);
 415             fiz1             = _mm_add_ps(fiz1,tz);
 416
 417             fjx3             = _mm_add_ps(fjx3,tx);
 418             fjy3             = _mm_add_ps(fjy3,ty);
 419             fjz3             = _mm_add_ps(fjz3,tz);
 420
 421             /**************************
 422              * CALCULATE INTERACTIONS *
 423              **************************/
 424
 425             /* COULOMB ELECTROSTATICS */
 426             velec            = _mm_mul_ps(qq21,rinv21);
 427             felec            = _mm_mul_ps(velec,rinvsq21);
 428
 429             /* Update potential sum for this i atom from the interaction with this j atom. */
 430             velecsum         = _mm_add_ps(velecsum,velec);
 431
 432             fscal            = felec;
 433
 434             /* Calculate temporary vectorial force */
 435             tx               = _mm_mul_ps(fscal,dx21);
 436             ty               = _mm_mul_ps(fscal,dy21);
 437             tz               = _mm_mul_ps(fscal,dz21);
 438
 439             /* Update vectorial force */
 440             fix2             = _mm_add_ps(fix2,tx);
 441             fiy2             = _mm_add_ps(fiy2,ty);
 442             fiz2             = _mm_add_ps(fiz2,tz);
 443
 444             fjx1             = _mm_add_ps(fjx1,tx);
 445             fjy1             = _mm_add_ps(fjy1,ty);
 446             fjz1             = _mm_add_ps(fjz1,tz);
 447
 448             /**************************
 449              * CALCULATE INTERACTIONS *
 450              **************************/
 451
 452             /* COULOMB ELECTROSTATICS */
 453             velec            = _mm_mul_ps(qq22,rinv22);
 454             felec            = _mm_mul_ps(velec,rinvsq22);
 455
 456             /* Update potential sum for this i atom from the interaction with this j atom. */
 457             velecsum         = _mm_add_ps(velecsum,velec);
 458
 459             fscal            = felec;
 460
 461             /* Calculate temporary vectorial force */
 462             tx               = _mm_mul_ps(fscal,dx22);
 463             ty               = _mm_mul_ps(fscal,dy22);
 464             tz               = _mm_mul_ps(fscal,dz22);
 465
 466             /* Update vectorial force */
 467             fix2             = _mm_add_ps(fix2,tx);
 468             fiy2             = _mm_add_ps(fiy2,ty);
 469             fiz2             = _mm_add_ps(fiz2,tz);
 470
 471             fjx2             = _mm_add_ps(fjx2,tx);
 472             fjy2             = _mm_add_ps(fjy2,ty);
 473             fjz2             = _mm_add_ps(fjz2,tz);
 474
 475             /**************************
 476              * CALCULATE INTERACTIONS *
 477              **************************/
 478
 479             /* COULOMB ELECTROSTATICS */
 480             velec            = _mm_mul_ps(qq23,rinv23);
 481             felec            = _mm_mul_ps(velec,rinvsq23);
 482
 483             /* Update potential sum for this i atom from the interaction with this j atom. */
 484             velecsum         = _mm_add_ps(velecsum,velec);
 485
 486             fscal            = felec;
 487
 488             /* Calculate temporary vectorial force */
 489             tx               = _mm_mul_ps(fscal,dx23);
 490             ty               = _mm_mul_ps(fscal,dy23);
 491             tz               = _mm_mul_ps(fscal,dz23);
 492
 493             /* Update vectorial force */
 494             fix2             = _mm_add_ps(fix2,tx);
 495             fiy2             = _mm_add_ps(fiy2,ty);
 496             fiz2             = _mm_add_ps(fiz2,tz);
 497
 498             fjx3             = _mm_add_ps(fjx3,tx);
 499             fjy3             = _mm_add_ps(fjy3,ty);
 500             fjz3             = _mm_add_ps(fjz3,tz);
 501
 502             /**************************
 503              * CALCULATE INTERACTIONS *
 504              **************************/
 505
 506             /* COULOMB ELECTROSTATICS */
 507             velec            = _mm_mul_ps(qq31,rinv31);
 508             felec            = _mm_mul_ps(velec,rinvsq31);
 509
 510             /* Update potential sum for this i atom from the interaction with this j atom. */
 511             velecsum         = _mm_add_ps(velecsum,velec);
 512
 513             fscal            = felec;
 514
 515             /* Calculate temporary vectorial force */
 516             tx               = _mm_mul_ps(fscal,dx31);
 517             ty               = _mm_mul_ps(fscal,dy31);
 518             tz               = _mm_mul_ps(fscal,dz31);
 519
 520             /* Update vectorial force */
 521             fix3             = _mm_add_ps(fix3,tx);
 522             fiy3             = _mm_add_ps(fiy3,ty);
 523             fiz3             = _mm_add_ps(fiz3,tz);
 524
 525             fjx1             = _mm_add_ps(fjx1,tx);
 526             fjy1             = _mm_add_ps(fjy1,ty);
 527             fjz1             = _mm_add_ps(fjz1,tz);
 528
 529             /**************************
 530              * CALCULATE INTERACTIONS *
 531              **************************/
 532
 533             /* COULOMB ELECTROSTATICS */
 534             velec            = _mm_mul_ps(qq32,rinv32);
 535             felec            = _mm_mul_ps(velec,rinvsq32);
 536
 537             /* Update potential sum for this i atom from the interaction with this j atom. */
 538             velecsum         = _mm_add_ps(velecsum,velec);
 539
 540             fscal            = felec;
 541
 542             /* Calculate temporary vectorial force */
 543             tx               = _mm_mul_ps(fscal,dx32);
 544             ty               = _mm_mul_ps(fscal,dy32);
 545             tz               = _mm_mul_ps(fscal,dz32);
 546
 547             /* Update vectorial force */
 548             fix3             = _mm_add_ps(fix3,tx);
 549             fiy3             = _mm_add_ps(fiy3,ty);
 550             fiz3             = _mm_add_ps(fiz3,tz);
 551
 552             fjx2             = _mm_add_ps(fjx2,tx);
 553             fjy2             = _mm_add_ps(fjy2,ty);
 554             fjz2             = _mm_add_ps(fjz2,tz);
 555
 556             /**************************
 557              * CALCULATE INTERACTIONS *
 558              **************************/
 559
 560             /* COULOMB ELECTROSTATICS */
 561             velec            = _mm_mul_ps(qq33,rinv33);
 562             felec            = _mm_mul_ps(velec,rinvsq33);
 563
 564             /* Update potential sum for this i atom from the interaction with this j atom. */
 565             velecsum         = _mm_add_ps(velecsum,velec);
 566
 567             fscal            = felec;
 568
 569             /* Calculate temporary vectorial force */
 570             tx               = _mm_mul_ps(fscal,dx33);
 571             ty               = _mm_mul_ps(fscal,dy33);
 572             tz               = _mm_mul_ps(fscal,dz33);
 573
 574             /* Update vectorial force */
 575             fix3             = _mm_add_ps(fix3,tx);
 576             fiy3             = _mm_add_ps(fiy3,ty);
 577             fiz3             = _mm_add_ps(fiz3,tz);
 578
 579             fjx3             = _mm_add_ps(fjx3,tx);
 580             fjy3             = _mm_add_ps(fjy3,ty);
 581             fjz3             = _mm_add_ps(fjz3,tz);
 582
 583             fjptrA             = f+j_coord_offsetA;
 584             fjptrB             = f+j_coord_offsetB;
 585             fjptrC             = f+j_coord_offsetC;
 586             fjptrD             = f+j_coord_offsetD;
 587
 588             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 589                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 590                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 591
 592             /* Inner loop uses 287 flops */
 593         }
 594
 595         if(jidx<j_index_end)
 596         {
 597
 598             /* Get j neighbor index, and coordinate index */
 599             jnrlistA         = jjnr[jidx];
 600             jnrlistB         = jjnr[jidx+1];
 601             jnrlistC         = jjnr[jidx+2];
 602             jnrlistD         = jjnr[jidx+3];
 603             /* Sign of each element will be negative for non-real atoms.
 604              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 605              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 606              */
 607             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 608             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 609             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 610             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 611             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 612             j_coord_offsetA  = DIM*jnrA;
 613             j_coord_offsetB  = DIM*jnrB;
 614             j_coord_offsetC  = DIM*jnrC;
 615             j_coord_offsetD  = DIM*jnrD;
 616
 617             /* load j atom coordinates */
 618             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 619                                               x+j_coord_offsetC,x+j_coord_offsetD,
 620                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 621                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 622
 623             /* Calculate displacement vector */
 624             dx00             = _mm_sub_ps(ix0,jx0);
 625             dy00             = _mm_sub_ps(iy0,jy0);
 626             dz00             = _mm_sub_ps(iz0,jz0);
 627             dx11             = _mm_sub_ps(ix1,jx1);
 628             dy11             = _mm_sub_ps(iy1,jy1);
 629             dz11             = _mm_sub_ps(iz1,jz1);
 630             dx12             = _mm_sub_ps(ix1,jx2);
 631             dy12             = _mm_sub_ps(iy1,jy2);
 632             dz12             = _mm_sub_ps(iz1,jz2);
 633             dx13             = _mm_sub_ps(ix1,jx3);
 634             dy13             = _mm_sub_ps(iy1,jy3);
 635             dz13             = _mm_sub_ps(iz1,jz3);
 636             dx21             = _mm_sub_ps(ix2,jx1);
 637             dy21             = _mm_sub_ps(iy2,jy1);
 638             dz21             = _mm_sub_ps(iz2,jz1);
 639             dx22             = _mm_sub_ps(ix2,jx2);
 640             dy22             = _mm_sub_ps(iy2,jy2);
 641             dz22             = _mm_sub_ps(iz2,jz2);
 642             dx23             = _mm_sub_ps(ix2,jx3);
 643             dy23             = _mm_sub_ps(iy2,jy3);
 644             dz23             = _mm_sub_ps(iz2,jz3);
 645             dx31             = _mm_sub_ps(ix3,jx1);
 646             dy31             = _mm_sub_ps(iy3,jy1);
 647             dz31             = _mm_sub_ps(iz3,jz1);
 648             dx32             = _mm_sub_ps(ix3,jx2);
 649             dy32             = _mm_sub_ps(iy3,jy2);
 650             dz32             = _mm_sub_ps(iz3,jz2);
 651             dx33             = _mm_sub_ps(ix3,jx3);
 652             dy33             = _mm_sub_ps(iy3,jy3);
 653             dz33             = _mm_sub_ps(iz3,jz3);
 654
 655             /* Calculate squared distance and things based on it */
 656             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 657             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 658             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 659             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 660             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 661             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 662             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 663             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 664             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 665             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 666
 667             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 668             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 669             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 670             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 671             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 672             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 673             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 674             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 675             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 676
 677             rinvsq00         = gmx_mm_inv_ps(rsq00);
 678             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 679             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 680             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 681             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 682             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 683             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 684             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 685             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 686             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 687
 688             fjx0             = _mm_setzero_ps();
 689             fjy0             = _mm_setzero_ps();
 690             fjz0             = _mm_setzero_ps();
 691             fjx1             = _mm_setzero_ps();
 692             fjy1             = _mm_setzero_ps();
 693             fjz1             = _mm_setzero_ps();
 694             fjx2             = _mm_setzero_ps();
 695             fjy2             = _mm_setzero_ps();
 696             fjz2             = _mm_setzero_ps();
 697             fjx3             = _mm_setzero_ps();
 698             fjy3             = _mm_setzero_ps();
 699             fjz3             = _mm_setzero_ps();
 700
 701             /**************************
 702              * CALCULATE INTERACTIONS *
 703              **************************/
 704
 705             /* LENNARD-JONES DISPERSION/REPULSION */
 706
 707             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 708             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 709             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 710             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 711             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 712
 713             /* Update potential sum for this i atom from the interaction with this j atom. */
 714             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 715             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 716
 717             fscal            = fvdw;
 718
 719             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 720
 721             /* Calculate temporary vectorial force */
 722             tx               = _mm_mul_ps(fscal,dx00);
 723             ty               = _mm_mul_ps(fscal,dy00);
 724             tz               = _mm_mul_ps(fscal,dz00);
 725
 726             /* Update vectorial force */
 727             fix0             = _mm_add_ps(fix0,tx);
 728             fiy0             = _mm_add_ps(fiy0,ty);
 729             fiz0             = _mm_add_ps(fiz0,tz);
 730
 731             fjx0             = _mm_add_ps(fjx0,tx);
 732             fjy0             = _mm_add_ps(fjy0,ty);
 733             fjz0             = _mm_add_ps(fjz0,tz);
 734
 735             /**************************
 736              * CALCULATE INTERACTIONS *
 737              **************************/
 738
 739             /* COULOMB ELECTROSTATICS */
 740             velec            = _mm_mul_ps(qq11,rinv11);
 741             felec            = _mm_mul_ps(velec,rinvsq11);
 742
 743             /* Update potential sum for this i atom from the interaction with this j atom. */
 744             velec            = _mm_andnot_ps(dummy_mask,velec);
 745             velecsum         = _mm_add_ps(velecsum,velec);
 746
 747             fscal            = felec;
 748
 749             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 750
 751             /* Calculate temporary vectorial force */
 752             tx               = _mm_mul_ps(fscal,dx11);
 753             ty               = _mm_mul_ps(fscal,dy11);
 754             tz               = _mm_mul_ps(fscal,dz11);
 755
 756             /* Update vectorial force */
 757             fix1             = _mm_add_ps(fix1,tx);
 758             fiy1             = _mm_add_ps(fiy1,ty);
 759             fiz1             = _mm_add_ps(fiz1,tz);
 760
 761             fjx1             = _mm_add_ps(fjx1,tx);
 762             fjy1             = _mm_add_ps(fjy1,ty);
 763             fjz1             = _mm_add_ps(fjz1,tz);
 764
 765             /**************************
 766              * CALCULATE INTERACTIONS *
 767              **************************/
 768
 769             /* COULOMB ELECTROSTATICS */
 770             velec            = _mm_mul_ps(qq12,rinv12);
 771             felec            = _mm_mul_ps(velec,rinvsq12);
 772
 773             /* Update potential sum for this i atom from the interaction with this j atom. */
 774             velec            = _mm_andnot_ps(dummy_mask,velec);
 775             velecsum         = _mm_add_ps(velecsum,velec);
 776
 777             fscal            = felec;
 778
 779             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 780
 781             /* Calculate temporary vectorial force */
 782             tx               = _mm_mul_ps(fscal,dx12);
 783             ty               = _mm_mul_ps(fscal,dy12);
 784             tz               = _mm_mul_ps(fscal,dz12);
 785
 786             /* Update vectorial force */
 787             fix1             = _mm_add_ps(fix1,tx);
 788             fiy1             = _mm_add_ps(fiy1,ty);
 789             fiz1             = _mm_add_ps(fiz1,tz);
 790
 791             fjx2             = _mm_add_ps(fjx2,tx);
 792             fjy2             = _mm_add_ps(fjy2,ty);
 793             fjz2             = _mm_add_ps(fjz2,tz);
 794
 795             /**************************
 796              * CALCULATE INTERACTIONS *
 797              **************************/
 798
 799             /* COULOMB ELECTROSTATICS */
 800             velec            = _mm_mul_ps(qq13,rinv13);
 801             felec            = _mm_mul_ps(velec,rinvsq13);
 802
 803             /* Update potential sum for this i atom from the interaction with this j atom. */
 804             velec            = _mm_andnot_ps(dummy_mask,velec);
 805             velecsum         = _mm_add_ps(velecsum,velec);
 806
 807             fscal            = felec;
 808
 809             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 810
 811             /* Calculate temporary vectorial force */
 812             tx               = _mm_mul_ps(fscal,dx13);
 813             ty               = _mm_mul_ps(fscal,dy13);
 814             tz               = _mm_mul_ps(fscal,dz13);
 815
 816             /* Update vectorial force */
 817             fix1             = _mm_add_ps(fix1,tx);
 818             fiy1             = _mm_add_ps(fiy1,ty);
 819             fiz1             = _mm_add_ps(fiz1,tz);
 820
 821             fjx3             = _mm_add_ps(fjx3,tx);
 822             fjy3             = _mm_add_ps(fjy3,ty);
 823             fjz3             = _mm_add_ps(fjz3,tz);
 824
 825             /**************************
 826              * CALCULATE INTERACTIONS *
 827              **************************/
 828
 829             /* COULOMB ELECTROSTATICS */
 830             velec            = _mm_mul_ps(qq21,rinv21);
 831             felec            = _mm_mul_ps(velec,rinvsq21);
 832
 833             /* Update potential sum for this i atom from the interaction with this j atom. */
 834             velec            = _mm_andnot_ps(dummy_mask,velec);
 835             velecsum         = _mm_add_ps(velecsum,velec);
 836
 837             fscal            = felec;
 838
 839             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 840
 841             /* Calculate temporary vectorial force */
 842             tx               = _mm_mul_ps(fscal,dx21);
 843             ty               = _mm_mul_ps(fscal,dy21);
 844             tz               = _mm_mul_ps(fscal,dz21);
 845
 846             /* Update vectorial force */
 847             fix2             = _mm_add_ps(fix2,tx);
 848             fiy2             = _mm_add_ps(fiy2,ty);
 849             fiz2             = _mm_add_ps(fiz2,tz);
 850
 851             fjx1             = _mm_add_ps(fjx1,tx);
 852             fjy1             = _mm_add_ps(fjy1,ty);
 853             fjz1             = _mm_add_ps(fjz1,tz);
 854
 855             /**************************
 856              * CALCULATE INTERACTIONS *
 857              **************************/
 858
 859             /* COULOMB ELECTROSTATICS */
 860             velec            = _mm_mul_ps(qq22,rinv22);
 861             felec            = _mm_mul_ps(velec,rinvsq22);
 862
 863             /* Update potential sum for this i atom from the interaction with this j atom. */
 864             velec            = _mm_andnot_ps(dummy_mask,velec);
 865             velecsum         = _mm_add_ps(velecsum,velec);
 866
 867             fscal            = felec;
 868
 869             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 870
 871             /* Calculate temporary vectorial force */
 872             tx               = _mm_mul_ps(fscal,dx22);
 873             ty               = _mm_mul_ps(fscal,dy22);
 874             tz               = _mm_mul_ps(fscal,dz22);
 875
 876             /* Update vectorial force */
 877             fix2             = _mm_add_ps(fix2,tx);
 878             fiy2             = _mm_add_ps(fiy2,ty);
 879             fiz2             = _mm_add_ps(fiz2,tz);
 880
 881             fjx2             = _mm_add_ps(fjx2,tx);
 882             fjy2             = _mm_add_ps(fjy2,ty);
 883             fjz2             = _mm_add_ps(fjz2,tz);
 884
 885             /**************************
 886              * CALCULATE INTERACTIONS *
 887              **************************/
 888
 889             /* COULOMB ELECTROSTATICS */
 890             velec            = _mm_mul_ps(qq23,rinv23);
 891             felec            = _mm_mul_ps(velec,rinvsq23);
 892
 893             /* Update potential sum for this i atom from the interaction with this j atom. */
 894             velec            = _mm_andnot_ps(dummy_mask,velec);
 895             velecsum         = _mm_add_ps(velecsum,velec);
 896
 897             fscal            = felec;
 898
 899             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 900
 901             /* Calculate temporary vectorial force */
 902             tx               = _mm_mul_ps(fscal,dx23);
 903             ty               = _mm_mul_ps(fscal,dy23);
 904             tz               = _mm_mul_ps(fscal,dz23);
 905
 906             /* Update vectorial force */
 907             fix2             = _mm_add_ps(fix2,tx);
 908             fiy2             = _mm_add_ps(fiy2,ty);
 909             fiz2             = _mm_add_ps(fiz2,tz);
 910
 911             fjx3             = _mm_add_ps(fjx3,tx);
 912             fjy3             = _mm_add_ps(fjy3,ty);
 913             fjz3             = _mm_add_ps(fjz3,tz);
 914
 915             /**************************
 916              * CALCULATE INTERACTIONS *
 917              **************************/
 918
 919             /* COULOMB ELECTROSTATICS */
 920             velec            = _mm_mul_ps(qq31,rinv31);
 921             felec            = _mm_mul_ps(velec,rinvsq31);
 922
 923             /* Update potential sum for this i atom from the interaction with this j atom. */
 924             velec            = _mm_andnot_ps(dummy_mask,velec);
 925             velecsum         = _mm_add_ps(velecsum,velec);
 926
 927             fscal            = felec;
 928
 929             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 930
 931             /* Calculate temporary vectorial force */
 932             tx               = _mm_mul_ps(fscal,dx31);
 933             ty               = _mm_mul_ps(fscal,dy31);
 934             tz               = _mm_mul_ps(fscal,dz31);
 935
 936             /* Update vectorial force */
 937             fix3             = _mm_add_ps(fix3,tx);
 938             fiy3             = _mm_add_ps(fiy3,ty);
 939             fiz3             = _mm_add_ps(fiz3,tz);
 940
 941             fjx1             = _mm_add_ps(fjx1,tx);
 942             fjy1             = _mm_add_ps(fjy1,ty);
 943             fjz1             = _mm_add_ps(fjz1,tz);
 944
 945             /**************************
 946              * CALCULATE INTERACTIONS *
 947              **************************/
 948
 949             /* COULOMB ELECTROSTATICS */
 950             velec            = _mm_mul_ps(qq32,rinv32);
 951             felec            = _mm_mul_ps(velec,rinvsq32);
 952
 953             /* Update potential sum for this i atom from the interaction with this j atom. */
 954             velec            = _mm_andnot_ps(dummy_mask,velec);
 955             velecsum         = _mm_add_ps(velecsum,velec);
 956
 957             fscal            = felec;
 958
 959             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 960
 961             /* Calculate temporary vectorial force */
 962             tx               = _mm_mul_ps(fscal,dx32);
 963             ty               = _mm_mul_ps(fscal,dy32);
 964             tz               = _mm_mul_ps(fscal,dz32);
 965
 966             /* Update vectorial force */
 967             fix3             = _mm_add_ps(fix3,tx);
 968             fiy3             = _mm_add_ps(fiy3,ty);
 969             fiz3             = _mm_add_ps(fiz3,tz);
 970
 971             fjx2             = _mm_add_ps(fjx2,tx);
 972             fjy2             = _mm_add_ps(fjy2,ty);
 973             fjz2             = _mm_add_ps(fjz2,tz);
 974
 975             /**************************
 976              * CALCULATE INTERACTIONS *
 977              **************************/
 978
 979             /* COULOMB ELECTROSTATICS */
 980             velec            = _mm_mul_ps(qq33,rinv33);
 981             felec            = _mm_mul_ps(velec,rinvsq33);
 982
 983             /* Update potential sum for this i atom from the interaction with this j atom. */
 984             velec            = _mm_andnot_ps(dummy_mask,velec);
 985             velecsum         = _mm_add_ps(velecsum,velec);
 986
 987             fscal            = felec;
 988
 989             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 990
 991             /* Calculate temporary vectorial force */
 992             tx               = _mm_mul_ps(fscal,dx33);
 993             ty               = _mm_mul_ps(fscal,dy33);
 994             tz               = _mm_mul_ps(fscal,dz33);
 995
 996             /* Update vectorial force */
 997             fix3             = _mm_add_ps(fix3,tx);
 998             fiy3             = _mm_add_ps(fiy3,ty);
 999             fiz3             = _mm_add_ps(fiz3,tz);
1000
1001             fjx3             = _mm_add_ps(fjx3,tx);
1002             fjy3             = _mm_add_ps(fjy3,ty);
1003             fjz3             = _mm_add_ps(fjz3,tz);
1004
1005             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1006             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1007             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1008             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1009
1010             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1011                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1012                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1013
1014             /* Inner loop uses 287 flops */
1015         }
1016
1017         /* End of innermost loop */
1018
1019         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1020                                               f+i_coord_offset,fshift+i_shift_offset);
1021
1022         ggid                        = gid[iidx];
1023         /* Update potential energies */
1024         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1025         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1026
1027         /* Increment number of inner iterations */
1028         inneriter                  += j_index_end - j_index_start;
1029
1030         /* Outer loop uses 26 flops */
1031     }
1032
1033     /* Increment number of outer iterations */
1034     outeriter        += nri;
1035
1036     /* Update outer/inner flops */
1037
1038     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*287);
1039 }
1040 /*
1041  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_sse4_1_single
1042  * Electrostatics interaction: Coulomb
1043  * VdW interaction:            LennardJones
1044  * Geometry:                   Water4-Water4
1045  * Calculate force/pot:        Force
1046  */
1047 void
1048 nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_sse4_1_single
1049                     (t_nblist                    * gmx_restrict       nlist,
1050                      rvec                        * gmx_restrict          xx,
1051                      rvec                        * gmx_restrict          ff,
1052                      t_forcerec                  * gmx_restrict          fr,
1053                      t_mdatoms                   * gmx_restrict     mdatoms,
1054                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1055                      t_nrnb                      * gmx_restrict        nrnb)
1056 {
1057     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1058      * just 0 for non-waters.
1059      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1060      * jnr indices corresponding to data put in the four positions in the SIMD register.
1061      */
1062     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1063     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1064     int              jnrA,jnrB,jnrC,jnrD;
1065     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1066     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1067     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1068     real             rcutoff_scalar;
1069     real             *shiftvec,*fshift,*x,*f;
1070     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1071     real             scratch[4*DIM];
1072     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1073     int              vdwioffset0;
1074     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1075     int              vdwioffset1;
1076     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1077     int              vdwioffset2;
1078     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1079     int              vdwioffset3;
1080     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1081     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1082     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1083     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1084     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1085     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1086     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1087     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1088     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1089     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1090     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1091     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1092     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1093     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1094     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1095     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1096     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1097     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1098     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1099     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1100     real             *charge;
1101     int              nvdwtype;
1102     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1103     int              *vdwtype;
1104     real             *vdwparam;
1105     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1106     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1107     __m128           dummy_mask,cutoff_mask;
1108     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1109     __m128           one     = _mm_set1_ps(1.0);
1110     __m128           two     = _mm_set1_ps(2.0);
1111     x                = xx[0];
1112     f                = ff[0];
1113
1114     nri              = nlist->nri;
1115     iinr             = nlist->iinr;
1116     jindex           = nlist->jindex;
1117     jjnr             = nlist->jjnr;
1118     shiftidx         = nlist->shift;
1119     gid              = nlist->gid;
1120     shiftvec         = fr->shift_vec[0];
1121     fshift           = fr->fshift[0];
1122     facel            = _mm_set1_ps(fr->epsfac);
1123     charge           = mdatoms->chargeA;
1124     nvdwtype         = fr->ntype;
1125     vdwparam         = fr->nbfp;
1126     vdwtype          = mdatoms->typeA;
1127
1128     /* Setup water-specific parameters */
1129     inr              = nlist->iinr[0];
1130     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1131     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1132     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1133     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1134
1135     jq1              = _mm_set1_ps(charge[inr+1]);
1136     jq2              = _mm_set1_ps(charge[inr+2]);
1137     jq3              = _mm_set1_ps(charge[inr+3]);
1138     vdwjidx0A        = 2*vdwtype[inr+0];
1139     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1140     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1141     qq11             = _mm_mul_ps(iq1,jq1);
1142     qq12             = _mm_mul_ps(iq1,jq2);
1143     qq13             = _mm_mul_ps(iq1,jq3);
1144     qq21             = _mm_mul_ps(iq2,jq1);
1145     qq22             = _mm_mul_ps(iq2,jq2);
1146     qq23             = _mm_mul_ps(iq2,jq3);
1147     qq31             = _mm_mul_ps(iq3,jq1);
1148     qq32             = _mm_mul_ps(iq3,jq2);
1149     qq33             = _mm_mul_ps(iq3,jq3);
1150
1151     /* Avoid stupid compiler warnings */
1152     jnrA = jnrB = jnrC = jnrD = 0;
1153     j_coord_offsetA = 0;
1154     j_coord_offsetB = 0;
1155     j_coord_offsetC = 0;
1156     j_coord_offsetD = 0;
1157
1158     outeriter        = 0;
1159     inneriter        = 0;
1160
1161     for(iidx=0;iidx<4*DIM;iidx++)
1162     {
1163         scratch[iidx] = 0.0;
1164     }
1165
1166     /* Start outer loop over neighborlists */
1167     for(iidx=0; iidx<nri; iidx++)
1168     {
1169         /* Load shift vector for this list */
1170         i_shift_offset   = DIM*shiftidx[iidx];
1171
1172         /* Load limits for loop over neighbors */
1173         j_index_start    = jindex[iidx];
1174         j_index_end      = jindex[iidx+1];
1175
1176         /* Get outer coordinate index */
1177         inr              = iinr[iidx];
1178         i_coord_offset   = DIM*inr;
1179
1180         /* Load i particle coords and add shift vector */
1181         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1182                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1183
1184         fix0             = _mm_setzero_ps();
1185         fiy0             = _mm_setzero_ps();
1186         fiz0             = _mm_setzero_ps();
1187         fix1             = _mm_setzero_ps();
1188         fiy1             = _mm_setzero_ps();
1189         fiz1             = _mm_setzero_ps();
1190         fix2             = _mm_setzero_ps();
1191         fiy2             = _mm_setzero_ps();
1192         fiz2             = _mm_setzero_ps();
1193         fix3             = _mm_setzero_ps();
1194         fiy3             = _mm_setzero_ps();
1195         fiz3             = _mm_setzero_ps();
1196
1197         /* Start inner kernel loop */
1198         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1199         {
1200
1201             /* Get j neighbor index, and coordinate index */
1202             jnrA             = jjnr[jidx];
1203             jnrB             = jjnr[jidx+1];
1204             jnrC             = jjnr[jidx+2];
1205             jnrD             = jjnr[jidx+3];
1206             j_coord_offsetA  = DIM*jnrA;
1207             j_coord_offsetB  = DIM*jnrB;
1208             j_coord_offsetC  = DIM*jnrC;
1209             j_coord_offsetD  = DIM*jnrD;
1210
1211             /* load j atom coordinates */
1212             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1213                                               x+j_coord_offsetC,x+j_coord_offsetD,
1214                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1215                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1216
1217             /* Calculate displacement vector */
1218             dx00             = _mm_sub_ps(ix0,jx0);
1219             dy00             = _mm_sub_ps(iy0,jy0);
1220             dz00             = _mm_sub_ps(iz0,jz0);
1221             dx11             = _mm_sub_ps(ix1,jx1);
1222             dy11             = _mm_sub_ps(iy1,jy1);
1223             dz11             = _mm_sub_ps(iz1,jz1);
1224             dx12             = _mm_sub_ps(ix1,jx2);
1225             dy12             = _mm_sub_ps(iy1,jy2);
1226             dz12             = _mm_sub_ps(iz1,jz2);
1227             dx13             = _mm_sub_ps(ix1,jx3);
1228             dy13             = _mm_sub_ps(iy1,jy3);
1229             dz13             = _mm_sub_ps(iz1,jz3);
1230             dx21             = _mm_sub_ps(ix2,jx1);
1231             dy21             = _mm_sub_ps(iy2,jy1);
1232             dz21             = _mm_sub_ps(iz2,jz1);
1233             dx22             = _mm_sub_ps(ix2,jx2);
1234             dy22             = _mm_sub_ps(iy2,jy2);
1235             dz22             = _mm_sub_ps(iz2,jz2);
1236             dx23             = _mm_sub_ps(ix2,jx3);
1237             dy23             = _mm_sub_ps(iy2,jy3);
1238             dz23             = _mm_sub_ps(iz2,jz3);
1239             dx31             = _mm_sub_ps(ix3,jx1);
1240             dy31             = _mm_sub_ps(iy3,jy1);
1241             dz31             = _mm_sub_ps(iz3,jz1);
1242             dx32             = _mm_sub_ps(ix3,jx2);
1243             dy32             = _mm_sub_ps(iy3,jy2);
1244             dz32             = _mm_sub_ps(iz3,jz2);
1245             dx33             = _mm_sub_ps(ix3,jx3);
1246             dy33             = _mm_sub_ps(iy3,jy3);
1247             dz33             = _mm_sub_ps(iz3,jz3);
1248
1249             /* Calculate squared distance and things based on it */
1250             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1251             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1252             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1253             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1254             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1255             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1256             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1257             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1258             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1259             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1260
1261             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1262             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1263             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1264             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1265             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1266             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1267             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1268             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1269             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1270
1271             rinvsq00         = gmx_mm_inv_ps(rsq00);
1272             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1273             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1274             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1275             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1276             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1277             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1278             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1279             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1280             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1281
1282             fjx0             = _mm_setzero_ps();
1283             fjy0             = _mm_setzero_ps();
1284             fjz0             = _mm_setzero_ps();
1285             fjx1             = _mm_setzero_ps();
1286             fjy1             = _mm_setzero_ps();
1287             fjz1             = _mm_setzero_ps();
1288             fjx2             = _mm_setzero_ps();
1289             fjy2             = _mm_setzero_ps();
1290             fjz2             = _mm_setzero_ps();
1291             fjx3             = _mm_setzero_ps();
1292             fjy3             = _mm_setzero_ps();
1293             fjz3             = _mm_setzero_ps();
1294
1295             /**************************
1296              * CALCULATE INTERACTIONS *
1297              **************************/
1298
1299             /* LENNARD-JONES DISPERSION/REPULSION */
1300
1301             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1302             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1303
1304             fscal            = fvdw;
1305
1306             /* Calculate temporary vectorial force */
1307             tx               = _mm_mul_ps(fscal,dx00);
1308             ty               = _mm_mul_ps(fscal,dy00);
1309             tz               = _mm_mul_ps(fscal,dz00);
1310
1311             /* Update vectorial force */
1312             fix0             = _mm_add_ps(fix0,tx);
1313             fiy0             = _mm_add_ps(fiy0,ty);
1314             fiz0             = _mm_add_ps(fiz0,tz);
1315
1316             fjx0             = _mm_add_ps(fjx0,tx);
1317             fjy0             = _mm_add_ps(fjy0,ty);
1318             fjz0             = _mm_add_ps(fjz0,tz);
1319
1320             /**************************
1321              * CALCULATE INTERACTIONS *
1322              **************************/
1323
1324             /* COULOMB ELECTROSTATICS */
1325             velec            = _mm_mul_ps(qq11,rinv11);
1326             felec            = _mm_mul_ps(velec,rinvsq11);
1327
1328             fscal            = felec;
1329
1330             /* Calculate temporary vectorial force */
1331             tx               = _mm_mul_ps(fscal,dx11);
1332             ty               = _mm_mul_ps(fscal,dy11);
1333             tz               = _mm_mul_ps(fscal,dz11);
1334
1335             /* Update vectorial force */
1336             fix1             = _mm_add_ps(fix1,tx);
1337             fiy1             = _mm_add_ps(fiy1,ty);
1338             fiz1             = _mm_add_ps(fiz1,tz);
1339
1340             fjx1             = _mm_add_ps(fjx1,tx);
1341             fjy1             = _mm_add_ps(fjy1,ty);
1342             fjz1             = _mm_add_ps(fjz1,tz);
1343
1344             /**************************
1345              * CALCULATE INTERACTIONS *
1346              **************************/
1347
1348             /* COULOMB ELECTROSTATICS */
1349             velec            = _mm_mul_ps(qq12,rinv12);
1350             felec            = _mm_mul_ps(velec,rinvsq12);
1351
1352             fscal            = felec;
1353
1354             /* Calculate temporary vectorial force */
1355             tx               = _mm_mul_ps(fscal,dx12);
1356             ty               = _mm_mul_ps(fscal,dy12);
1357             tz               = _mm_mul_ps(fscal,dz12);
1358
1359             /* Update vectorial force */
1360             fix1             = _mm_add_ps(fix1,tx);
1361             fiy1             = _mm_add_ps(fiy1,ty);
1362             fiz1             = _mm_add_ps(fiz1,tz);
1363
1364             fjx2             = _mm_add_ps(fjx2,tx);
1365             fjy2             = _mm_add_ps(fjy2,ty);
1366             fjz2             = _mm_add_ps(fjz2,tz);
1367
1368             /**************************
1369              * CALCULATE INTERACTIONS *
1370              **************************/
1371
1372             /* COULOMB ELECTROSTATICS */
1373             velec            = _mm_mul_ps(qq13,rinv13);
1374             felec            = _mm_mul_ps(velec,rinvsq13);
1375
1376             fscal            = felec;
1377
1378             /* Calculate temporary vectorial force */
1379             tx               = _mm_mul_ps(fscal,dx13);
1380             ty               = _mm_mul_ps(fscal,dy13);
1381             tz               = _mm_mul_ps(fscal,dz13);
1382
1383             /* Update vectorial force */
1384             fix1             = _mm_add_ps(fix1,tx);
1385             fiy1             = _mm_add_ps(fiy1,ty);
1386             fiz1             = _mm_add_ps(fiz1,tz);
1387
1388             fjx3             = _mm_add_ps(fjx3,tx);
1389             fjy3             = _mm_add_ps(fjy3,ty);
1390             fjz3             = _mm_add_ps(fjz3,tz);
1391
1392             /**************************
1393              * CALCULATE INTERACTIONS *
1394              **************************/
1395
1396             /* COULOMB ELECTROSTATICS */
1397             velec            = _mm_mul_ps(qq21,rinv21);
1398             felec            = _mm_mul_ps(velec,rinvsq21);
1399
1400             fscal            = felec;
1401
1402             /* Calculate temporary vectorial force */
1403             tx               = _mm_mul_ps(fscal,dx21);
1404             ty               = _mm_mul_ps(fscal,dy21);
1405             tz               = _mm_mul_ps(fscal,dz21);
1406
1407             /* Update vectorial force */
1408             fix2             = _mm_add_ps(fix2,tx);
1409             fiy2             = _mm_add_ps(fiy2,ty);
1410             fiz2             = _mm_add_ps(fiz2,tz);
1411
1412             fjx1             = _mm_add_ps(fjx1,tx);
1413             fjy1             = _mm_add_ps(fjy1,ty);
1414             fjz1             = _mm_add_ps(fjz1,tz);
1415
1416             /**************************
1417              * CALCULATE INTERACTIONS *
1418              **************************/
1419
1420             /* COULOMB ELECTROSTATICS */
1421             velec            = _mm_mul_ps(qq22,rinv22);
1422             felec            = _mm_mul_ps(velec,rinvsq22);
1423
1424             fscal            = felec;
1425
1426             /* Calculate temporary vectorial force */
1427             tx               = _mm_mul_ps(fscal,dx22);
1428             ty               = _mm_mul_ps(fscal,dy22);
1429             tz               = _mm_mul_ps(fscal,dz22);
1430
1431             /* Update vectorial force */
1432             fix2             = _mm_add_ps(fix2,tx);
1433             fiy2             = _mm_add_ps(fiy2,ty);
1434             fiz2             = _mm_add_ps(fiz2,tz);
1435
1436             fjx2             = _mm_add_ps(fjx2,tx);
1437             fjy2             = _mm_add_ps(fjy2,ty);
1438             fjz2             = _mm_add_ps(fjz2,tz);
1439
1440             /**************************
1441              * CALCULATE INTERACTIONS *
1442              **************************/
1443
1444             /* COULOMB ELECTROSTATICS */
1445             velec            = _mm_mul_ps(qq23,rinv23);
1446             felec            = _mm_mul_ps(velec,rinvsq23);
1447
1448             fscal            = felec;
1449
1450             /* Calculate temporary vectorial force */
1451             tx               = _mm_mul_ps(fscal,dx23);
1452             ty               = _mm_mul_ps(fscal,dy23);
1453             tz               = _mm_mul_ps(fscal,dz23);
1454
1455             /* Update vectorial force */
1456             fix2             = _mm_add_ps(fix2,tx);
1457             fiy2             = _mm_add_ps(fiy2,ty);
1458             fiz2             = _mm_add_ps(fiz2,tz);
1459
1460             fjx3             = _mm_add_ps(fjx3,tx);
1461             fjy3             = _mm_add_ps(fjy3,ty);
1462             fjz3             = _mm_add_ps(fjz3,tz);
1463
1464             /**************************
1465              * CALCULATE INTERACTIONS *
1466              **************************/
1467
1468             /* COULOMB ELECTROSTATICS */
1469             velec            = _mm_mul_ps(qq31,rinv31);
1470             felec            = _mm_mul_ps(velec,rinvsq31);
1471
1472             fscal            = felec;
1473
1474             /* Calculate temporary vectorial force */
1475             tx               = _mm_mul_ps(fscal,dx31);
1476             ty               = _mm_mul_ps(fscal,dy31);
1477             tz               = _mm_mul_ps(fscal,dz31);
1478
1479             /* Update vectorial force */
1480             fix3             = _mm_add_ps(fix3,tx);
1481             fiy3             = _mm_add_ps(fiy3,ty);
1482             fiz3             = _mm_add_ps(fiz3,tz);
1483
1484             fjx1             = _mm_add_ps(fjx1,tx);
1485             fjy1             = _mm_add_ps(fjy1,ty);
1486             fjz1             = _mm_add_ps(fjz1,tz);
1487
1488             /**************************
1489              * CALCULATE INTERACTIONS *
1490              **************************/
1491
1492             /* COULOMB ELECTROSTATICS */
1493             velec            = _mm_mul_ps(qq32,rinv32);
1494             felec            = _mm_mul_ps(velec,rinvsq32);
1495
1496             fscal            = felec;
1497
1498             /* Calculate temporary vectorial force */
1499             tx               = _mm_mul_ps(fscal,dx32);
1500             ty               = _mm_mul_ps(fscal,dy32);
1501             tz               = _mm_mul_ps(fscal,dz32);
1502
1503             /* Update vectorial force */
1504             fix3             = _mm_add_ps(fix3,tx);
1505             fiy3             = _mm_add_ps(fiy3,ty);
1506             fiz3             = _mm_add_ps(fiz3,tz);
1507
1508             fjx2             = _mm_add_ps(fjx2,tx);
1509             fjy2             = _mm_add_ps(fjy2,ty);
1510             fjz2             = _mm_add_ps(fjz2,tz);
1511
1512             /**************************
1513              * CALCULATE INTERACTIONS *
1514              **************************/
1515
1516             /* COULOMB ELECTROSTATICS */
1517             velec            = _mm_mul_ps(qq33,rinv33);
1518             felec            = _mm_mul_ps(velec,rinvsq33);
1519
1520             fscal            = felec;
1521
1522             /* Calculate temporary vectorial force */
1523             tx               = _mm_mul_ps(fscal,dx33);
1524             ty               = _mm_mul_ps(fscal,dy33);
1525             tz               = _mm_mul_ps(fscal,dz33);
1526
1527             /* Update vectorial force */
1528             fix3             = _mm_add_ps(fix3,tx);
1529             fiy3             = _mm_add_ps(fiy3,ty);
1530             fiz3             = _mm_add_ps(fiz3,tz);
1531
1532             fjx3             = _mm_add_ps(fjx3,tx);
1533             fjy3             = _mm_add_ps(fjy3,ty);
1534             fjz3             = _mm_add_ps(fjz3,tz);
1535
1536             fjptrA             = f+j_coord_offsetA;
1537             fjptrB             = f+j_coord_offsetB;
1538             fjptrC             = f+j_coord_offsetC;
1539             fjptrD             = f+j_coord_offsetD;
1540
1541             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1542                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1543                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1544
1545             /* Inner loop uses 273 flops */
1546         }
1547
1548         if(jidx<j_index_end)
1549         {
1550
1551             /* Get j neighbor index, and coordinate index */
1552             jnrlistA         = jjnr[jidx];
1553             jnrlistB         = jjnr[jidx+1];
1554             jnrlistC         = jjnr[jidx+2];
1555             jnrlistD         = jjnr[jidx+3];
1556             /* Sign of each element will be negative for non-real atoms.
1557              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1558              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1559              */
1560             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1561             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1562             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1563             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1564             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1565             j_coord_offsetA  = DIM*jnrA;
1566             j_coord_offsetB  = DIM*jnrB;
1567             j_coord_offsetC  = DIM*jnrC;
1568             j_coord_offsetD  = DIM*jnrD;
1569
1570             /* load j atom coordinates */
1571             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1572                                               x+j_coord_offsetC,x+j_coord_offsetD,
1573                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1574                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1575
1576             /* Calculate displacement vector */
1577             dx00             = _mm_sub_ps(ix0,jx0);
1578             dy00             = _mm_sub_ps(iy0,jy0);
1579             dz00             = _mm_sub_ps(iz0,jz0);
1580             dx11             = _mm_sub_ps(ix1,jx1);
1581             dy11             = _mm_sub_ps(iy1,jy1);
1582             dz11             = _mm_sub_ps(iz1,jz1);
1583             dx12             = _mm_sub_ps(ix1,jx2);
1584             dy12             = _mm_sub_ps(iy1,jy2);
1585             dz12             = _mm_sub_ps(iz1,jz2);
1586             dx13             = _mm_sub_ps(ix1,jx3);
1587             dy13             = _mm_sub_ps(iy1,jy3);
1588             dz13             = _mm_sub_ps(iz1,jz3);
1589             dx21             = _mm_sub_ps(ix2,jx1);
1590             dy21             = _mm_sub_ps(iy2,jy1);
1591             dz21             = _mm_sub_ps(iz2,jz1);
1592             dx22             = _mm_sub_ps(ix2,jx2);
1593             dy22             = _mm_sub_ps(iy2,jy2);
1594             dz22             = _mm_sub_ps(iz2,jz2);
1595             dx23             = _mm_sub_ps(ix2,jx3);
1596             dy23             = _mm_sub_ps(iy2,jy3);
1597             dz23             = _mm_sub_ps(iz2,jz3);
1598             dx31             = _mm_sub_ps(ix3,jx1);
1599             dy31             = _mm_sub_ps(iy3,jy1);
1600             dz31             = _mm_sub_ps(iz3,jz1);
1601             dx32             = _mm_sub_ps(ix3,jx2);
1602             dy32             = _mm_sub_ps(iy3,jy2);
1603             dz32             = _mm_sub_ps(iz3,jz2);
1604             dx33             = _mm_sub_ps(ix3,jx3);
1605             dy33             = _mm_sub_ps(iy3,jy3);
1606             dz33             = _mm_sub_ps(iz3,jz3);
1607
1608             /* Calculate squared distance and things based on it */
1609             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1610             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1611             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1612             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1613             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1614             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1615             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1616             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1617             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1618             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1619
1620             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1621             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1622             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1623             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1624             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1625             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1626             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1627             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1628             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1629
1630             rinvsq00         = gmx_mm_inv_ps(rsq00);
1631             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1632             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1633             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1634             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1635             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1636             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1637             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1638             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1639             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1640
1641             fjx0             = _mm_setzero_ps();
1642             fjy0             = _mm_setzero_ps();
1643             fjz0             = _mm_setzero_ps();
1644             fjx1             = _mm_setzero_ps();
1645             fjy1             = _mm_setzero_ps();
1646             fjz1             = _mm_setzero_ps();
1647             fjx2             = _mm_setzero_ps();
1648             fjy2             = _mm_setzero_ps();
1649             fjz2             = _mm_setzero_ps();
1650             fjx3             = _mm_setzero_ps();
1651             fjy3             = _mm_setzero_ps();
1652             fjz3             = _mm_setzero_ps();
1653
1654             /**************************
1655              * CALCULATE INTERACTIONS *
1656              **************************/
1657
1658             /* LENNARD-JONES DISPERSION/REPULSION */
1659
1660             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1661             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1662
1663             fscal            = fvdw;
1664
1665             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1666
1667             /* Calculate temporary vectorial force */
1668             tx               = _mm_mul_ps(fscal,dx00);
1669             ty               = _mm_mul_ps(fscal,dy00);
1670             tz               = _mm_mul_ps(fscal,dz00);
1671
1672             /* Update vectorial force */
1673             fix0             = _mm_add_ps(fix0,tx);
1674             fiy0             = _mm_add_ps(fiy0,ty);
1675             fiz0             = _mm_add_ps(fiz0,tz);
1676
1677             fjx0             = _mm_add_ps(fjx0,tx);
1678             fjy0             = _mm_add_ps(fjy0,ty);
1679             fjz0             = _mm_add_ps(fjz0,tz);
1680
1681             /**************************
1682              * CALCULATE INTERACTIONS *
1683              **************************/
1684
1685             /* COULOMB ELECTROSTATICS */
1686             velec            = _mm_mul_ps(qq11,rinv11);
1687             felec            = _mm_mul_ps(velec,rinvsq11);
1688
1689             fscal            = felec;
1690
1691             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1692
1693             /* Calculate temporary vectorial force */
1694             tx               = _mm_mul_ps(fscal,dx11);
1695             ty               = _mm_mul_ps(fscal,dy11);
1696             tz               = _mm_mul_ps(fscal,dz11);
1697
1698             /* Update vectorial force */
1699             fix1             = _mm_add_ps(fix1,tx);
1700             fiy1             = _mm_add_ps(fiy1,ty);
1701             fiz1             = _mm_add_ps(fiz1,tz);
1702
1703             fjx1             = _mm_add_ps(fjx1,tx);
1704             fjy1             = _mm_add_ps(fjy1,ty);
1705             fjz1             = _mm_add_ps(fjz1,tz);
1706
1707             /**************************
1708              * CALCULATE INTERACTIONS *
1709              **************************/
1710
1711             /* COULOMB ELECTROSTATICS */
1712             velec            = _mm_mul_ps(qq12,rinv12);
1713             felec            = _mm_mul_ps(velec,rinvsq12);
1714
1715             fscal            = felec;
1716
1717             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1718
1719             /* Calculate temporary vectorial force */
1720             tx               = _mm_mul_ps(fscal,dx12);
1721             ty               = _mm_mul_ps(fscal,dy12);
1722             tz               = _mm_mul_ps(fscal,dz12);
1723
1724             /* Update vectorial force */
1725             fix1             = _mm_add_ps(fix1,tx);
1726             fiy1             = _mm_add_ps(fiy1,ty);
1727             fiz1             = _mm_add_ps(fiz1,tz);
1728
1729             fjx2             = _mm_add_ps(fjx2,tx);
1730             fjy2             = _mm_add_ps(fjy2,ty);
1731             fjz2             = _mm_add_ps(fjz2,tz);
1732
1733             /**************************
1734              * CALCULATE INTERACTIONS *
1735              **************************/
1736
1737             /* COULOMB ELECTROSTATICS */
1738             velec            = _mm_mul_ps(qq13,rinv13);
1739             felec            = _mm_mul_ps(velec,rinvsq13);
1740
1741             fscal            = felec;
1742
1743             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1744
1745             /* Calculate temporary vectorial force */
1746             tx               = _mm_mul_ps(fscal,dx13);
1747             ty               = _mm_mul_ps(fscal,dy13);
1748             tz               = _mm_mul_ps(fscal,dz13);
1749
1750             /* Update vectorial force */
1751             fix1             = _mm_add_ps(fix1,tx);
1752             fiy1             = _mm_add_ps(fiy1,ty);
1753             fiz1             = _mm_add_ps(fiz1,tz);
1754
1755             fjx3             = _mm_add_ps(fjx3,tx);
1756             fjy3             = _mm_add_ps(fjy3,ty);
1757             fjz3             = _mm_add_ps(fjz3,tz);
1758
1759             /**************************
1760              * CALCULATE INTERACTIONS *
1761              **************************/
1762
1763             /* COULOMB ELECTROSTATICS */
1764             velec            = _mm_mul_ps(qq21,rinv21);
1765             felec            = _mm_mul_ps(velec,rinvsq21);
1766
1767             fscal            = felec;
1768
1769             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1770
1771             /* Calculate temporary vectorial force */
1772             tx               = _mm_mul_ps(fscal,dx21);
1773             ty               = _mm_mul_ps(fscal,dy21);
1774             tz               = _mm_mul_ps(fscal,dz21);
1775
1776             /* Update vectorial force */
1777             fix2             = _mm_add_ps(fix2,tx);
1778             fiy2             = _mm_add_ps(fiy2,ty);
1779             fiz2             = _mm_add_ps(fiz2,tz);
1780
1781             fjx1             = _mm_add_ps(fjx1,tx);
1782             fjy1             = _mm_add_ps(fjy1,ty);
1783             fjz1             = _mm_add_ps(fjz1,tz);
1784
1785             /**************************
1786              * CALCULATE INTERACTIONS *
1787              **************************/
1788
1789             /* COULOMB ELECTROSTATICS */
1790             velec            = _mm_mul_ps(qq22,rinv22);
1791             felec            = _mm_mul_ps(velec,rinvsq22);
1792
1793             fscal            = felec;
1794
1795             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1796
1797             /* Calculate temporary vectorial force */
1798             tx               = _mm_mul_ps(fscal,dx22);
1799             ty               = _mm_mul_ps(fscal,dy22);
1800             tz               = _mm_mul_ps(fscal,dz22);
1801
1802             /* Update vectorial force */
1803             fix2             = _mm_add_ps(fix2,tx);
1804             fiy2             = _mm_add_ps(fiy2,ty);
1805             fiz2             = _mm_add_ps(fiz2,tz);
1806
1807             fjx2             = _mm_add_ps(fjx2,tx);
1808             fjy2             = _mm_add_ps(fjy2,ty);
1809             fjz2             = _mm_add_ps(fjz2,tz);
1810
1811             /**************************
1812              * CALCULATE INTERACTIONS *
1813              **************************/
1814
1815             /* COULOMB ELECTROSTATICS */
1816             velec            = _mm_mul_ps(qq23,rinv23);
1817             felec            = _mm_mul_ps(velec,rinvsq23);
1818
1819             fscal            = felec;
1820
1821             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1822
1823             /* Calculate temporary vectorial force */
1824             tx               = _mm_mul_ps(fscal,dx23);
1825             ty               = _mm_mul_ps(fscal,dy23);
1826             tz               = _mm_mul_ps(fscal,dz23);
1827
1828             /* Update vectorial force */
1829             fix2             = _mm_add_ps(fix2,tx);
1830             fiy2             = _mm_add_ps(fiy2,ty);
1831             fiz2             = _mm_add_ps(fiz2,tz);
1832
1833             fjx3             = _mm_add_ps(fjx3,tx);
1834             fjy3             = _mm_add_ps(fjy3,ty);
1835             fjz3             = _mm_add_ps(fjz3,tz);
1836
1837             /**************************
1838              * CALCULATE INTERACTIONS *
1839              **************************/
1840
1841             /* COULOMB ELECTROSTATICS */
1842             velec            = _mm_mul_ps(qq31,rinv31);
1843             felec            = _mm_mul_ps(velec,rinvsq31);
1844
1845             fscal            = felec;
1846
1847             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1848
1849             /* Calculate temporary vectorial force */
1850             tx               = _mm_mul_ps(fscal,dx31);
1851             ty               = _mm_mul_ps(fscal,dy31);
1852             tz               = _mm_mul_ps(fscal,dz31);
1853
1854             /* Update vectorial force */
1855             fix3             = _mm_add_ps(fix3,tx);
1856             fiy3             = _mm_add_ps(fiy3,ty);
1857             fiz3             = _mm_add_ps(fiz3,tz);
1858
1859             fjx1             = _mm_add_ps(fjx1,tx);
1860             fjy1             = _mm_add_ps(fjy1,ty);
1861             fjz1             = _mm_add_ps(fjz1,tz);
1862
1863             /**************************
1864              * CALCULATE INTERACTIONS *
1865              **************************/
1866
1867             /* COULOMB ELECTROSTATICS */
1868             velec            = _mm_mul_ps(qq32,rinv32);
1869             felec            = _mm_mul_ps(velec,rinvsq32);
1870
1871             fscal            = felec;
1872
1873             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1874
1875             /* Calculate temporary vectorial force */
1876             tx               = _mm_mul_ps(fscal,dx32);
1877             ty               = _mm_mul_ps(fscal,dy32);
1878             tz               = _mm_mul_ps(fscal,dz32);
1879
1880             /* Update vectorial force */
1881             fix3             = _mm_add_ps(fix3,tx);
1882             fiy3             = _mm_add_ps(fiy3,ty);
1883             fiz3             = _mm_add_ps(fiz3,tz);
1884
1885             fjx2             = _mm_add_ps(fjx2,tx);
1886             fjy2             = _mm_add_ps(fjy2,ty);
1887             fjz2             = _mm_add_ps(fjz2,tz);
1888
1889             /**************************
1890              * CALCULATE INTERACTIONS *
1891              **************************/
1892
1893             /* COULOMB ELECTROSTATICS */
1894             velec            = _mm_mul_ps(qq33,rinv33);
1895             felec            = _mm_mul_ps(velec,rinvsq33);
1896
1897             fscal            = felec;
1898
1899             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1900
1901             /* Calculate temporary vectorial force */
1902             tx               = _mm_mul_ps(fscal,dx33);
1903             ty               = _mm_mul_ps(fscal,dy33);
1904             tz               = _mm_mul_ps(fscal,dz33);
1905
1906             /* Update vectorial force */
1907             fix3             = _mm_add_ps(fix3,tx);
1908             fiy3             = _mm_add_ps(fiy3,ty);
1909             fiz3             = _mm_add_ps(fiz3,tz);
1910
1911             fjx3             = _mm_add_ps(fjx3,tx);
1912             fjy3             = _mm_add_ps(fjy3,ty);
1913             fjz3             = _mm_add_ps(fjz3,tz);
1914
1915             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1916             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1917             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1918             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1919
1920             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1921                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1922                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1923
1924             /* Inner loop uses 273 flops */
1925         }
1926
1927         /* End of innermost loop */
1928
1929         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1930                                               f+i_coord_offset,fshift+i_shift_offset);
1931
1932         /* Increment number of inner iterations */
1933         inneriter                  += j_index_end - j_index_start;
1934
1935         /* Outer loop uses 24 flops */
1936     }
1937
1938     /* Increment number of outer iterations */
1939     outeriter        += nri;
1940
1941     /* Update outer/inner flops */
1942
1943     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*273);
1944 }