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) 2012,2013,2014,2015,2017, 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 "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_sse4_1_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4W4_VF_sse4_1_single
  51  * Electrostatics interaction: Coulomb
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Water4-Water4
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecCoul_VdwLJ_GeomW4W4_VF_sse4_1_single
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  75     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  80     real             scratch[4*DIM];
  81     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwioffset1;
  85     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  86     int              vdwioffset2;
  87     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  88     int              vdwioffset3;
  89     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  90     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  91     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  92     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  93     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  94     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  95     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  96     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  97     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  98     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  99     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 100     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 101     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 102     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 103     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 104     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 105     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 106     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 107     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 108     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 109     real             *charge;
 110     int              nvdwtype;
 111     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 112     int              *vdwtype;
 113     real             *vdwparam;
 114     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 115     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 116     __m128           dummy_mask,cutoff_mask;
 117     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 118     __m128           one     = _mm_set1_ps(1.0);
 119     __m128           two     = _mm_set1_ps(2.0);
 120     x                = xx[0];
 121     f                = ff[0];
 122
 123     nri              = nlist->nri;
 124     iinr             = nlist->iinr;
 125     jindex           = nlist->jindex;
 126     jjnr             = nlist->jjnr;
 127     shiftidx         = nlist->shift;
 128     gid              = nlist->gid;
 129     shiftvec         = fr->shift_vec[0];
 130     fshift           = fr->fshift[0];
 131     facel            = _mm_set1_ps(fr->ic->epsfac);
 132     charge           = mdatoms->chargeA;
 133     nvdwtype         = fr->ntype;
 134     vdwparam         = fr->nbfp;
 135     vdwtype          = mdatoms->typeA;
 136
 137     /* Setup water-specific parameters */
 138     inr              = nlist->iinr[0];
 139     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 140     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 141     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 142     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 143
 144     jq1              = _mm_set1_ps(charge[inr+1]);
 145     jq2              = _mm_set1_ps(charge[inr+2]);
 146     jq3              = _mm_set1_ps(charge[inr+3]);
 147     vdwjidx0A        = 2*vdwtype[inr+0];
 148     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 149     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 150     qq11             = _mm_mul_ps(iq1,jq1);
 151     qq12             = _mm_mul_ps(iq1,jq2);
 152     qq13             = _mm_mul_ps(iq1,jq3);
 153     qq21             = _mm_mul_ps(iq2,jq1);
 154     qq22             = _mm_mul_ps(iq2,jq2);
 155     qq23             = _mm_mul_ps(iq2,jq3);
 156     qq31             = _mm_mul_ps(iq3,jq1);
 157     qq32             = _mm_mul_ps(iq3,jq2);
 158     qq33             = _mm_mul_ps(iq3,jq3);
 159
 160     /* Avoid stupid compiler warnings */
 161     jnrA = jnrB = jnrC = jnrD = 0;
 162     j_coord_offsetA = 0;
 163     j_coord_offsetB = 0;
 164     j_coord_offsetC = 0;
 165     j_coord_offsetD = 0;
 166
 167     outeriter        = 0;
 168     inneriter        = 0;
 169
 170     for(iidx=0;iidx<4*DIM;iidx++)
 171     {
 172         scratch[iidx] = 0.0;
 173     }
 174
 175     /* Start outer loop over neighborlists */
 176     for(iidx=0; iidx<nri; iidx++)
 177     {
 178         /* Load shift vector for this list */
 179         i_shift_offset   = DIM*shiftidx[iidx];
 180
 181         /* Load limits for loop over neighbors */
 182         j_index_start    = jindex[iidx];
 183         j_index_end      = jindex[iidx+1];
 184
 185         /* Get outer coordinate index */
 186         inr              = iinr[iidx];
 187         i_coord_offset   = DIM*inr;
 188
 189         /* Load i particle coords and add shift vector */
 190         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 191                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 192
 193         fix0             = _mm_setzero_ps();
 194         fiy0             = _mm_setzero_ps();
 195         fiz0             = _mm_setzero_ps();
 196         fix1             = _mm_setzero_ps();
 197         fiy1             = _mm_setzero_ps();
 198         fiz1             = _mm_setzero_ps();
 199         fix2             = _mm_setzero_ps();
 200         fiy2             = _mm_setzero_ps();
 201         fiz2             = _mm_setzero_ps();
 202         fix3             = _mm_setzero_ps();
 203         fiy3             = _mm_setzero_ps();
 204         fiz3             = _mm_setzero_ps();
 205
 206         /* Reset potential sums */
 207         velecsum         = _mm_setzero_ps();
 208         vvdwsum          = _mm_setzero_ps();
 209
 210         /* Start inner kernel loop */
 211         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 212         {
 213
 214             /* Get j neighbor index, and coordinate index */
 215             jnrA             = jjnr[jidx];
 216             jnrB             = jjnr[jidx+1];
 217             jnrC             = jjnr[jidx+2];
 218             jnrD             = jjnr[jidx+3];
 219             j_coord_offsetA  = DIM*jnrA;
 220             j_coord_offsetB  = DIM*jnrB;
 221             j_coord_offsetC  = DIM*jnrC;
 222             j_coord_offsetD  = DIM*jnrD;
 223
 224             /* load j atom coordinates */
 225             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 226                                               x+j_coord_offsetC,x+j_coord_offsetD,
 227                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 228                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 229
 230             /* Calculate displacement vector */
 231             dx00             = _mm_sub_ps(ix0,jx0);
 232             dy00             = _mm_sub_ps(iy0,jy0);
 233             dz00             = _mm_sub_ps(iz0,jz0);
 234             dx11             = _mm_sub_ps(ix1,jx1);
 235             dy11             = _mm_sub_ps(iy1,jy1);
 236             dz11             = _mm_sub_ps(iz1,jz1);
 237             dx12             = _mm_sub_ps(ix1,jx2);
 238             dy12             = _mm_sub_ps(iy1,jy2);
 239             dz12             = _mm_sub_ps(iz1,jz2);
 240             dx13             = _mm_sub_ps(ix1,jx3);
 241             dy13             = _mm_sub_ps(iy1,jy3);
 242             dz13             = _mm_sub_ps(iz1,jz3);
 243             dx21             = _mm_sub_ps(ix2,jx1);
 244             dy21             = _mm_sub_ps(iy2,jy1);
 245             dz21             = _mm_sub_ps(iz2,jz1);
 246             dx22             = _mm_sub_ps(ix2,jx2);
 247             dy22             = _mm_sub_ps(iy2,jy2);
 248             dz22             = _mm_sub_ps(iz2,jz2);
 249             dx23             = _mm_sub_ps(ix2,jx3);
 250             dy23             = _mm_sub_ps(iy2,jy3);
 251             dz23             = _mm_sub_ps(iz2,jz3);
 252             dx31             = _mm_sub_ps(ix3,jx1);
 253             dy31             = _mm_sub_ps(iy3,jy1);
 254             dz31             = _mm_sub_ps(iz3,jz1);
 255             dx32             = _mm_sub_ps(ix3,jx2);
 256             dy32             = _mm_sub_ps(iy3,jy2);
 257             dz32             = _mm_sub_ps(iz3,jz2);
 258             dx33             = _mm_sub_ps(ix3,jx3);
 259             dy33             = _mm_sub_ps(iy3,jy3);
 260             dz33             = _mm_sub_ps(iz3,jz3);
 261
 262             /* Calculate squared distance and things based on it */
 263             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 264             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 265             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 266             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 267             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 268             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 269             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 270             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 271             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 272             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 273
 274             rinv11           = sse41_invsqrt_f(rsq11);
 275             rinv12           = sse41_invsqrt_f(rsq12);
 276             rinv13           = sse41_invsqrt_f(rsq13);
 277             rinv21           = sse41_invsqrt_f(rsq21);
 278             rinv22           = sse41_invsqrt_f(rsq22);
 279             rinv23           = sse41_invsqrt_f(rsq23);
 280             rinv31           = sse41_invsqrt_f(rsq31);
 281             rinv32           = sse41_invsqrt_f(rsq32);
 282             rinv33           = sse41_invsqrt_f(rsq33);
 283
 284             rinvsq00         = sse41_inv_f(rsq00);
 285             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 286             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 287             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 288             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 289             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 290             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 291             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 292             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 293             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 294
 295             fjx0             = _mm_setzero_ps();
 296             fjy0             = _mm_setzero_ps();
 297             fjz0             = _mm_setzero_ps();
 298             fjx1             = _mm_setzero_ps();
 299             fjy1             = _mm_setzero_ps();
 300             fjz1             = _mm_setzero_ps();
 301             fjx2             = _mm_setzero_ps();
 302             fjy2             = _mm_setzero_ps();
 303             fjz2             = _mm_setzero_ps();
 304             fjx3             = _mm_setzero_ps();
 305             fjy3             = _mm_setzero_ps();
 306             fjz3             = _mm_setzero_ps();
 307
 308             /**************************
 309              * CALCULATE INTERACTIONS *
 310              **************************/
 311
 312             /* LENNARD-JONES DISPERSION/REPULSION */
 313
 314             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 315             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 316             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 317             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 318             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 319
 320             /* Update potential sum for this i atom from the interaction with this j atom. */
 321             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 322
 323             fscal            = fvdw;
 324
 325             /* Calculate temporary vectorial force */
 326             tx               = _mm_mul_ps(fscal,dx00);
 327             ty               = _mm_mul_ps(fscal,dy00);
 328             tz               = _mm_mul_ps(fscal,dz00);
 329
 330             /* Update vectorial force */
 331             fix0             = _mm_add_ps(fix0,tx);
 332             fiy0             = _mm_add_ps(fiy0,ty);
 333             fiz0             = _mm_add_ps(fiz0,tz);
 334
 335             fjx0             = _mm_add_ps(fjx0,tx);
 336             fjy0             = _mm_add_ps(fjy0,ty);
 337             fjz0             = _mm_add_ps(fjz0,tz);
 338
 339             /**************************
 340              * CALCULATE INTERACTIONS *
 341              **************************/
 342
 343             /* COULOMB ELECTROSTATICS */
 344             velec            = _mm_mul_ps(qq11,rinv11);
 345             felec            = _mm_mul_ps(velec,rinvsq11);
 346
 347             /* Update potential sum for this i atom from the interaction with this j atom. */
 348             velecsum         = _mm_add_ps(velecsum,velec);
 349
 350             fscal            = felec;
 351
 352             /* Calculate temporary vectorial force */
 353             tx               = _mm_mul_ps(fscal,dx11);
 354             ty               = _mm_mul_ps(fscal,dy11);
 355             tz               = _mm_mul_ps(fscal,dz11);
 356
 357             /* Update vectorial force */
 358             fix1             = _mm_add_ps(fix1,tx);
 359             fiy1             = _mm_add_ps(fiy1,ty);
 360             fiz1             = _mm_add_ps(fiz1,tz);
 361
 362             fjx1             = _mm_add_ps(fjx1,tx);
 363             fjy1             = _mm_add_ps(fjy1,ty);
 364             fjz1             = _mm_add_ps(fjz1,tz);
 365
 366             /**************************
 367              * CALCULATE INTERACTIONS *
 368              **************************/
 369
 370             /* COULOMB ELECTROSTATICS */
 371             velec            = _mm_mul_ps(qq12,rinv12);
 372             felec            = _mm_mul_ps(velec,rinvsq12);
 373
 374             /* Update potential sum for this i atom from the interaction with this j atom. */
 375             velecsum         = _mm_add_ps(velecsum,velec);
 376
 377             fscal            = felec;
 378
 379             /* Calculate temporary vectorial force */
 380             tx               = _mm_mul_ps(fscal,dx12);
 381             ty               = _mm_mul_ps(fscal,dy12);
 382             tz               = _mm_mul_ps(fscal,dz12);
 383
 384             /* Update vectorial force */
 385             fix1             = _mm_add_ps(fix1,tx);
 386             fiy1             = _mm_add_ps(fiy1,ty);
 387             fiz1             = _mm_add_ps(fiz1,tz);
 388
 389             fjx2             = _mm_add_ps(fjx2,tx);
 390             fjy2             = _mm_add_ps(fjy2,ty);
 391             fjz2             = _mm_add_ps(fjz2,tz);
 392
 393             /**************************
 394              * CALCULATE INTERACTIONS *
 395              **************************/
 396
 397             /* COULOMB ELECTROSTATICS */
 398             velec            = _mm_mul_ps(qq13,rinv13);
 399             felec            = _mm_mul_ps(velec,rinvsq13);
 400
 401             /* Update potential sum for this i atom from the interaction with this j atom. */
 402             velecsum         = _mm_add_ps(velecsum,velec);
 403
 404             fscal            = felec;
 405
 406             /* Calculate temporary vectorial force */
 407             tx               = _mm_mul_ps(fscal,dx13);
 408             ty               = _mm_mul_ps(fscal,dy13);
 409             tz               = _mm_mul_ps(fscal,dz13);
 410
 411             /* Update vectorial force */
 412             fix1             = _mm_add_ps(fix1,tx);
 413             fiy1             = _mm_add_ps(fiy1,ty);
 414             fiz1             = _mm_add_ps(fiz1,tz);
 415
 416             fjx3             = _mm_add_ps(fjx3,tx);
 417             fjy3             = _mm_add_ps(fjy3,ty);
 418             fjz3             = _mm_add_ps(fjz3,tz);
 419
 420             /**************************
 421              * CALCULATE INTERACTIONS *
 422              **************************/
 423
 424             /* COULOMB ELECTROSTATICS */
 425             velec            = _mm_mul_ps(qq21,rinv21);
 426             felec            = _mm_mul_ps(velec,rinvsq21);
 427
 428             /* Update potential sum for this i atom from the interaction with this j atom. */
 429             velecsum         = _mm_add_ps(velecsum,velec);
 430
 431             fscal            = felec;
 432
 433             /* Calculate temporary vectorial force */
 434             tx               = _mm_mul_ps(fscal,dx21);
 435             ty               = _mm_mul_ps(fscal,dy21);
 436             tz               = _mm_mul_ps(fscal,dz21);
 437
 438             /* Update vectorial force */
 439             fix2             = _mm_add_ps(fix2,tx);
 440             fiy2             = _mm_add_ps(fiy2,ty);
 441             fiz2             = _mm_add_ps(fiz2,tz);
 442
 443             fjx1             = _mm_add_ps(fjx1,tx);
 444             fjy1             = _mm_add_ps(fjy1,ty);
 445             fjz1             = _mm_add_ps(fjz1,tz);
 446
 447             /**************************
 448              * CALCULATE INTERACTIONS *
 449              **************************/
 450
 451             /* COULOMB ELECTROSTATICS */
 452             velec            = _mm_mul_ps(qq22,rinv22);
 453             felec            = _mm_mul_ps(velec,rinvsq22);
 454
 455             /* Update potential sum for this i atom from the interaction with this j atom. */
 456             velecsum         = _mm_add_ps(velecsum,velec);
 457
 458             fscal            = felec;
 459
 460             /* Calculate temporary vectorial force */
 461             tx               = _mm_mul_ps(fscal,dx22);
 462             ty               = _mm_mul_ps(fscal,dy22);
 463             tz               = _mm_mul_ps(fscal,dz22);
 464
 465             /* Update vectorial force */
 466             fix2             = _mm_add_ps(fix2,tx);
 467             fiy2             = _mm_add_ps(fiy2,ty);
 468             fiz2             = _mm_add_ps(fiz2,tz);
 469
 470             fjx2             = _mm_add_ps(fjx2,tx);
 471             fjy2             = _mm_add_ps(fjy2,ty);
 472             fjz2             = _mm_add_ps(fjz2,tz);
 473
 474             /**************************
 475              * CALCULATE INTERACTIONS *
 476              **************************/
 477
 478             /* COULOMB ELECTROSTATICS */
 479             velec            = _mm_mul_ps(qq23,rinv23);
 480             felec            = _mm_mul_ps(velec,rinvsq23);
 481
 482             /* Update potential sum for this i atom from the interaction with this j atom. */
 483             velecsum         = _mm_add_ps(velecsum,velec);
 484
 485             fscal            = felec;
 486
 487             /* Calculate temporary vectorial force */
 488             tx               = _mm_mul_ps(fscal,dx23);
 489             ty               = _mm_mul_ps(fscal,dy23);
 490             tz               = _mm_mul_ps(fscal,dz23);
 491
 492             /* Update vectorial force */
 493             fix2             = _mm_add_ps(fix2,tx);
 494             fiy2             = _mm_add_ps(fiy2,ty);
 495             fiz2             = _mm_add_ps(fiz2,tz);
 496
 497             fjx3             = _mm_add_ps(fjx3,tx);
 498             fjy3             = _mm_add_ps(fjy3,ty);
 499             fjz3             = _mm_add_ps(fjz3,tz);
 500
 501             /**************************
 502              * CALCULATE INTERACTIONS *
 503              **************************/
 504
 505             /* COULOMB ELECTROSTATICS */
 506             velec            = _mm_mul_ps(qq31,rinv31);
 507             felec            = _mm_mul_ps(velec,rinvsq31);
 508
 509             /* Update potential sum for this i atom from the interaction with this j atom. */
 510             velecsum         = _mm_add_ps(velecsum,velec);
 511
 512             fscal            = felec;
 513
 514             /* Calculate temporary vectorial force */
 515             tx               = _mm_mul_ps(fscal,dx31);
 516             ty               = _mm_mul_ps(fscal,dy31);
 517             tz               = _mm_mul_ps(fscal,dz31);
 518
 519             /* Update vectorial force */
 520             fix3             = _mm_add_ps(fix3,tx);
 521             fiy3             = _mm_add_ps(fiy3,ty);
 522             fiz3             = _mm_add_ps(fiz3,tz);
 523
 524             fjx1             = _mm_add_ps(fjx1,tx);
 525             fjy1             = _mm_add_ps(fjy1,ty);
 526             fjz1             = _mm_add_ps(fjz1,tz);
 527
 528             /**************************
 529              * CALCULATE INTERACTIONS *
 530              **************************/
 531
 532             /* COULOMB ELECTROSTATICS */
 533             velec            = _mm_mul_ps(qq32,rinv32);
 534             felec            = _mm_mul_ps(velec,rinvsq32);
 535
 536             /* Update potential sum for this i atom from the interaction with this j atom. */
 537             velecsum         = _mm_add_ps(velecsum,velec);
 538
 539             fscal            = felec;
 540
 541             /* Calculate temporary vectorial force */
 542             tx               = _mm_mul_ps(fscal,dx32);
 543             ty               = _mm_mul_ps(fscal,dy32);
 544             tz               = _mm_mul_ps(fscal,dz32);
 545
 546             /* Update vectorial force */
 547             fix3             = _mm_add_ps(fix3,tx);
 548             fiy3             = _mm_add_ps(fiy3,ty);
 549             fiz3             = _mm_add_ps(fiz3,tz);
 550
 551             fjx2             = _mm_add_ps(fjx2,tx);
 552             fjy2             = _mm_add_ps(fjy2,ty);
 553             fjz2             = _mm_add_ps(fjz2,tz);
 554
 555             /**************************
 556              * CALCULATE INTERACTIONS *
 557              **************************/
 558
 559             /* COULOMB ELECTROSTATICS */
 560             velec            = _mm_mul_ps(qq33,rinv33);
 561             felec            = _mm_mul_ps(velec,rinvsq33);
 562
 563             /* Update potential sum for this i atom from the interaction with this j atom. */
 564             velecsum         = _mm_add_ps(velecsum,velec);
 565
 566             fscal            = felec;
 567
 568             /* Calculate temporary vectorial force */
 569             tx               = _mm_mul_ps(fscal,dx33);
 570             ty               = _mm_mul_ps(fscal,dy33);
 571             tz               = _mm_mul_ps(fscal,dz33);
 572
 573             /* Update vectorial force */
 574             fix3             = _mm_add_ps(fix3,tx);
 575             fiy3             = _mm_add_ps(fiy3,ty);
 576             fiz3             = _mm_add_ps(fiz3,tz);
 577
 578             fjx3             = _mm_add_ps(fjx3,tx);
 579             fjy3             = _mm_add_ps(fjy3,ty);
 580             fjz3             = _mm_add_ps(fjz3,tz);
 581
 582             fjptrA             = f+j_coord_offsetA;
 583             fjptrB             = f+j_coord_offsetB;
 584             fjptrC             = f+j_coord_offsetC;
 585             fjptrD             = f+j_coord_offsetD;
 586
 587             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 588                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 589                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 590
 591             /* Inner loop uses 287 flops */
 592         }
 593
 594         if(jidx<j_index_end)
 595         {
 596
 597             /* Get j neighbor index, and coordinate index */
 598             jnrlistA         = jjnr[jidx];
 599             jnrlistB         = jjnr[jidx+1];
 600             jnrlistC         = jjnr[jidx+2];
 601             jnrlistD         = jjnr[jidx+3];
 602             /* Sign of each element will be negative for non-real atoms.
 603              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 604              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 605              */
 606             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 607             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 608             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 609             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 610             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 611             j_coord_offsetA  = DIM*jnrA;
 612             j_coord_offsetB  = DIM*jnrB;
 613             j_coord_offsetC  = DIM*jnrC;
 614             j_coord_offsetD  = DIM*jnrD;
 615
 616             /* load j atom coordinates */
 617             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 618                                               x+j_coord_offsetC,x+j_coord_offsetD,
 619                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 620                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 621
 622             /* Calculate displacement vector */
 623             dx00             = _mm_sub_ps(ix0,jx0);
 624             dy00             = _mm_sub_ps(iy0,jy0);
 625             dz00             = _mm_sub_ps(iz0,jz0);
 626             dx11             = _mm_sub_ps(ix1,jx1);
 627             dy11             = _mm_sub_ps(iy1,jy1);
 628             dz11             = _mm_sub_ps(iz1,jz1);
 629             dx12             = _mm_sub_ps(ix1,jx2);
 630             dy12             = _mm_sub_ps(iy1,jy2);
 631             dz12             = _mm_sub_ps(iz1,jz2);
 632             dx13             = _mm_sub_ps(ix1,jx3);
 633             dy13             = _mm_sub_ps(iy1,jy3);
 634             dz13             = _mm_sub_ps(iz1,jz3);
 635             dx21             = _mm_sub_ps(ix2,jx1);
 636             dy21             = _mm_sub_ps(iy2,jy1);
 637             dz21             = _mm_sub_ps(iz2,jz1);
 638             dx22             = _mm_sub_ps(ix2,jx2);
 639             dy22             = _mm_sub_ps(iy2,jy2);
 640             dz22             = _mm_sub_ps(iz2,jz2);
 641             dx23             = _mm_sub_ps(ix2,jx3);
 642             dy23             = _mm_sub_ps(iy2,jy3);
 643             dz23             = _mm_sub_ps(iz2,jz3);
 644             dx31             = _mm_sub_ps(ix3,jx1);
 645             dy31             = _mm_sub_ps(iy3,jy1);
 646             dz31             = _mm_sub_ps(iz3,jz1);
 647             dx32             = _mm_sub_ps(ix3,jx2);
 648             dy32             = _mm_sub_ps(iy3,jy2);
 649             dz32             = _mm_sub_ps(iz3,jz2);
 650             dx33             = _mm_sub_ps(ix3,jx3);
 651             dy33             = _mm_sub_ps(iy3,jy3);
 652             dz33             = _mm_sub_ps(iz3,jz3);
 653
 654             /* Calculate squared distance and things based on it */
 655             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 656             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 657             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 658             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 659             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 660             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 661             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 662             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 663             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 664             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 665
 666             rinv11           = sse41_invsqrt_f(rsq11);
 667             rinv12           = sse41_invsqrt_f(rsq12);
 668             rinv13           = sse41_invsqrt_f(rsq13);
 669             rinv21           = sse41_invsqrt_f(rsq21);
 670             rinv22           = sse41_invsqrt_f(rsq22);
 671             rinv23           = sse41_invsqrt_f(rsq23);
 672             rinv31           = sse41_invsqrt_f(rsq31);
 673             rinv32           = sse41_invsqrt_f(rsq32);
 674             rinv33           = sse41_invsqrt_f(rsq33);
 675
 676             rinvsq00         = sse41_inv_f(rsq00);
 677             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 678             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 679             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 680             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 681             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 682             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 683             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 684             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 685             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 686
 687             fjx0             = _mm_setzero_ps();
 688             fjy0             = _mm_setzero_ps();
 689             fjz0             = _mm_setzero_ps();
 690             fjx1             = _mm_setzero_ps();
 691             fjy1             = _mm_setzero_ps();
 692             fjz1             = _mm_setzero_ps();
 693             fjx2             = _mm_setzero_ps();
 694             fjy2             = _mm_setzero_ps();
 695             fjz2             = _mm_setzero_ps();
 696             fjx3             = _mm_setzero_ps();
 697             fjy3             = _mm_setzero_ps();
 698             fjz3             = _mm_setzero_ps();
 699
 700             /**************************
 701              * CALCULATE INTERACTIONS *
 702              **************************/
 703
 704             /* LENNARD-JONES DISPERSION/REPULSION */
 705
 706             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 707             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 708             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 709             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 710             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 711
 712             /* Update potential sum for this i atom from the interaction with this j atom. */
 713             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 714             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 715
 716             fscal            = fvdw;
 717
 718             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 719
 720             /* Calculate temporary vectorial force */
 721             tx               = _mm_mul_ps(fscal,dx00);
 722             ty               = _mm_mul_ps(fscal,dy00);
 723             tz               = _mm_mul_ps(fscal,dz00);
 724
 725             /* Update vectorial force */
 726             fix0             = _mm_add_ps(fix0,tx);
 727             fiy0             = _mm_add_ps(fiy0,ty);
 728             fiz0             = _mm_add_ps(fiz0,tz);
 729
 730             fjx0             = _mm_add_ps(fjx0,tx);
 731             fjy0             = _mm_add_ps(fjy0,ty);
 732             fjz0             = _mm_add_ps(fjz0,tz);
 733
 734             /**************************
 735              * CALCULATE INTERACTIONS *
 736              **************************/
 737
 738             /* COULOMB ELECTROSTATICS */
 739             velec            = _mm_mul_ps(qq11,rinv11);
 740             felec            = _mm_mul_ps(velec,rinvsq11);
 741
 742             /* Update potential sum for this i atom from the interaction with this j atom. */
 743             velec            = _mm_andnot_ps(dummy_mask,velec);
 744             velecsum         = _mm_add_ps(velecsum,velec);
 745
 746             fscal            = felec;
 747
 748             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 749
 750             /* Calculate temporary vectorial force */
 751             tx               = _mm_mul_ps(fscal,dx11);
 752             ty               = _mm_mul_ps(fscal,dy11);
 753             tz               = _mm_mul_ps(fscal,dz11);
 754
 755             /* Update vectorial force */
 756             fix1             = _mm_add_ps(fix1,tx);
 757             fiy1             = _mm_add_ps(fiy1,ty);
 758             fiz1             = _mm_add_ps(fiz1,tz);
 759
 760             fjx1             = _mm_add_ps(fjx1,tx);
 761             fjy1             = _mm_add_ps(fjy1,ty);
 762             fjz1             = _mm_add_ps(fjz1,tz);
 763
 764             /**************************
 765              * CALCULATE INTERACTIONS *
 766              **************************/
 767
 768             /* COULOMB ELECTROSTATICS */
 769             velec            = _mm_mul_ps(qq12,rinv12);
 770             felec            = _mm_mul_ps(velec,rinvsq12);
 771
 772             /* Update potential sum for this i atom from the interaction with this j atom. */
 773             velec            = _mm_andnot_ps(dummy_mask,velec);
 774             velecsum         = _mm_add_ps(velecsum,velec);
 775
 776             fscal            = felec;
 777
 778             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 779
 780             /* Calculate temporary vectorial force */
 781             tx               = _mm_mul_ps(fscal,dx12);
 782             ty               = _mm_mul_ps(fscal,dy12);
 783             tz               = _mm_mul_ps(fscal,dz12);
 784
 785             /* Update vectorial force */
 786             fix1             = _mm_add_ps(fix1,tx);
 787             fiy1             = _mm_add_ps(fiy1,ty);
 788             fiz1             = _mm_add_ps(fiz1,tz);
 789
 790             fjx2             = _mm_add_ps(fjx2,tx);
 791             fjy2             = _mm_add_ps(fjy2,ty);
 792             fjz2             = _mm_add_ps(fjz2,tz);
 793
 794             /**************************
 795              * CALCULATE INTERACTIONS *
 796              **************************/
 797
 798             /* COULOMB ELECTROSTATICS */
 799             velec            = _mm_mul_ps(qq13,rinv13);
 800             felec            = _mm_mul_ps(velec,rinvsq13);
 801
 802             /* Update potential sum for this i atom from the interaction with this j atom. */
 803             velec            = _mm_andnot_ps(dummy_mask,velec);
 804             velecsum         = _mm_add_ps(velecsum,velec);
 805
 806             fscal            = felec;
 807
 808             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 809
 810             /* Calculate temporary vectorial force */
 811             tx               = _mm_mul_ps(fscal,dx13);
 812             ty               = _mm_mul_ps(fscal,dy13);
 813             tz               = _mm_mul_ps(fscal,dz13);
 814
 815             /* Update vectorial force */
 816             fix1             = _mm_add_ps(fix1,tx);
 817             fiy1             = _mm_add_ps(fiy1,ty);
 818             fiz1             = _mm_add_ps(fiz1,tz);
 819
 820             fjx3             = _mm_add_ps(fjx3,tx);
 821             fjy3             = _mm_add_ps(fjy3,ty);
 822             fjz3             = _mm_add_ps(fjz3,tz);
 823
 824             /**************************
 825              * CALCULATE INTERACTIONS *
 826              **************************/
 827
 828             /* COULOMB ELECTROSTATICS */
 829             velec            = _mm_mul_ps(qq21,rinv21);
 830             felec            = _mm_mul_ps(velec,rinvsq21);
 831
 832             /* Update potential sum for this i atom from the interaction with this j atom. */
 833             velec            = _mm_andnot_ps(dummy_mask,velec);
 834             velecsum         = _mm_add_ps(velecsum,velec);
 835
 836             fscal            = felec;
 837
 838             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 839
 840             /* Calculate temporary vectorial force */
 841             tx               = _mm_mul_ps(fscal,dx21);
 842             ty               = _mm_mul_ps(fscal,dy21);
 843             tz               = _mm_mul_ps(fscal,dz21);
 844
 845             /* Update vectorial force */
 846             fix2             = _mm_add_ps(fix2,tx);
 847             fiy2             = _mm_add_ps(fiy2,ty);
 848             fiz2             = _mm_add_ps(fiz2,tz);
 849
 850             fjx1             = _mm_add_ps(fjx1,tx);
 851             fjy1             = _mm_add_ps(fjy1,ty);
 852             fjz1             = _mm_add_ps(fjz1,tz);
 853
 854             /**************************
 855              * CALCULATE INTERACTIONS *
 856              **************************/
 857
 858             /* COULOMB ELECTROSTATICS */
 859             velec            = _mm_mul_ps(qq22,rinv22);
 860             felec            = _mm_mul_ps(velec,rinvsq22);
 861
 862             /* Update potential sum for this i atom from the interaction with this j atom. */
 863             velec            = _mm_andnot_ps(dummy_mask,velec);
 864             velecsum         = _mm_add_ps(velecsum,velec);
 865
 866             fscal            = felec;
 867
 868             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 869
 870             /* Calculate temporary vectorial force */
 871             tx               = _mm_mul_ps(fscal,dx22);
 872             ty               = _mm_mul_ps(fscal,dy22);
 873             tz               = _mm_mul_ps(fscal,dz22);
 874
 875             /* Update vectorial force */
 876             fix2             = _mm_add_ps(fix2,tx);
 877             fiy2             = _mm_add_ps(fiy2,ty);
 878             fiz2             = _mm_add_ps(fiz2,tz);
 879
 880             fjx2             = _mm_add_ps(fjx2,tx);
 881             fjy2             = _mm_add_ps(fjy2,ty);
 882             fjz2             = _mm_add_ps(fjz2,tz);
 883
 884             /**************************
 885              * CALCULATE INTERACTIONS *
 886              **************************/
 887
 888             /* COULOMB ELECTROSTATICS */
 889             velec            = _mm_mul_ps(qq23,rinv23);
 890             felec            = _mm_mul_ps(velec,rinvsq23);
 891
 892             /* Update potential sum for this i atom from the interaction with this j atom. */
 893             velec            = _mm_andnot_ps(dummy_mask,velec);
 894             velecsum         = _mm_add_ps(velecsum,velec);
 895
 896             fscal            = felec;
 897
 898             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 899
 900             /* Calculate temporary vectorial force */
 901             tx               = _mm_mul_ps(fscal,dx23);
 902             ty               = _mm_mul_ps(fscal,dy23);
 903             tz               = _mm_mul_ps(fscal,dz23);
 904
 905             /* Update vectorial force */
 906             fix2             = _mm_add_ps(fix2,tx);
 907             fiy2             = _mm_add_ps(fiy2,ty);
 908             fiz2             = _mm_add_ps(fiz2,tz);
 909
 910             fjx3             = _mm_add_ps(fjx3,tx);
 911             fjy3             = _mm_add_ps(fjy3,ty);
 912             fjz3             = _mm_add_ps(fjz3,tz);
 913
 914             /**************************
 915              * CALCULATE INTERACTIONS *
 916              **************************/
 917
 918             /* COULOMB ELECTROSTATICS */
 919             velec            = _mm_mul_ps(qq31,rinv31);
 920             felec            = _mm_mul_ps(velec,rinvsq31);
 921
 922             /* Update potential sum for this i atom from the interaction with this j atom. */
 923             velec            = _mm_andnot_ps(dummy_mask,velec);
 924             velecsum         = _mm_add_ps(velecsum,velec);
 925
 926             fscal            = felec;
 927
 928             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 929
 930             /* Calculate temporary vectorial force */
 931             tx               = _mm_mul_ps(fscal,dx31);
 932             ty               = _mm_mul_ps(fscal,dy31);
 933             tz               = _mm_mul_ps(fscal,dz31);
 934
 935             /* Update vectorial force */
 936             fix3             = _mm_add_ps(fix3,tx);
 937             fiy3             = _mm_add_ps(fiy3,ty);
 938             fiz3             = _mm_add_ps(fiz3,tz);
 939
 940             fjx1             = _mm_add_ps(fjx1,tx);
 941             fjy1             = _mm_add_ps(fjy1,ty);
 942             fjz1             = _mm_add_ps(fjz1,tz);
 943
 944             /**************************
 945              * CALCULATE INTERACTIONS *
 946              **************************/
 947
 948             /* COULOMB ELECTROSTATICS */
 949             velec            = _mm_mul_ps(qq32,rinv32);
 950             felec            = _mm_mul_ps(velec,rinvsq32);
 951
 952             /* Update potential sum for this i atom from the interaction with this j atom. */
 953             velec            = _mm_andnot_ps(dummy_mask,velec);
 954             velecsum         = _mm_add_ps(velecsum,velec);
 955
 956             fscal            = felec;
 957
 958             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 959
 960             /* Calculate temporary vectorial force */
 961             tx               = _mm_mul_ps(fscal,dx32);
 962             ty               = _mm_mul_ps(fscal,dy32);
 963             tz               = _mm_mul_ps(fscal,dz32);
 964
 965             /* Update vectorial force */
 966             fix3             = _mm_add_ps(fix3,tx);
 967             fiy3             = _mm_add_ps(fiy3,ty);
 968             fiz3             = _mm_add_ps(fiz3,tz);
 969
 970             fjx2             = _mm_add_ps(fjx2,tx);
 971             fjy2             = _mm_add_ps(fjy2,ty);
 972             fjz2             = _mm_add_ps(fjz2,tz);
 973
 974             /**************************
 975              * CALCULATE INTERACTIONS *
 976              **************************/
 977
 978             /* COULOMB ELECTROSTATICS */
 979             velec            = _mm_mul_ps(qq33,rinv33);
 980             felec            = _mm_mul_ps(velec,rinvsq33);
 981
 982             /* Update potential sum for this i atom from the interaction with this j atom. */
 983             velec            = _mm_andnot_ps(dummy_mask,velec);
 984             velecsum         = _mm_add_ps(velecsum,velec);
 985
 986             fscal            = felec;
 987
 988             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 989
 990             /* Calculate temporary vectorial force */
 991             tx               = _mm_mul_ps(fscal,dx33);
 992             ty               = _mm_mul_ps(fscal,dy33);
 993             tz               = _mm_mul_ps(fscal,dz33);
 994
 995             /* Update vectorial force */
 996             fix3             = _mm_add_ps(fix3,tx);
 997             fiy3             = _mm_add_ps(fiy3,ty);
 998             fiz3             = _mm_add_ps(fiz3,tz);
 999
1000             fjx3             = _mm_add_ps(fjx3,tx);
1001             fjy3             = _mm_add_ps(fjy3,ty);
1002             fjz3             = _mm_add_ps(fjz3,tz);
1003
1004             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1005             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1006             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1007             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1008
1009             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1010                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1011                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1012
1013             /* Inner loop uses 287 flops */
1014         }
1015
1016         /* End of innermost loop */
1017
1018         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1019                                               f+i_coord_offset,fshift+i_shift_offset);
1020
1021         ggid                        = gid[iidx];
1022         /* Update potential energies */
1023         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1024         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1025
1026         /* Increment number of inner iterations */
1027         inneriter                  += j_index_end - j_index_start;
1028
1029         /* Outer loop uses 26 flops */
1030     }
1031
1032     /* Increment number of outer iterations */
1033     outeriter        += nri;
1034
1035     /* Update outer/inner flops */
1036
1037     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*287);
1038 }
1039 /*
1040  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_sse4_1_single
1041  * Electrostatics interaction: Coulomb
1042  * VdW interaction:            LennardJones
1043  * Geometry:                   Water4-Water4
1044  * Calculate force/pot:        Force
1045  */
1046 void
1047 nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_sse4_1_single
1048                     (t_nblist                    * gmx_restrict       nlist,
1049                      rvec                        * gmx_restrict          xx,
1050                      rvec                        * gmx_restrict          ff,
1051                      struct t_forcerec           * gmx_restrict          fr,
1052                      t_mdatoms                   * gmx_restrict     mdatoms,
1053                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1054                      t_nrnb                      * gmx_restrict        nrnb)
1055 {
1056     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1057      * just 0 for non-waters.
1058      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1059      * jnr indices corresponding to data put in the four positions in the SIMD register.
1060      */
1061     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1062     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1063     int              jnrA,jnrB,jnrC,jnrD;
1064     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1065     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1066     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1067     real             rcutoff_scalar;
1068     real             *shiftvec,*fshift,*x,*f;
1069     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1070     real             scratch[4*DIM];
1071     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1072     int              vdwioffset0;
1073     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1074     int              vdwioffset1;
1075     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1076     int              vdwioffset2;
1077     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1078     int              vdwioffset3;
1079     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1080     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1081     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1082     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1083     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1084     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1085     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1086     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1087     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1088     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1089     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1090     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1091     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1092     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1093     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1094     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1095     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1096     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1097     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1098     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1099     real             *charge;
1100     int              nvdwtype;
1101     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1102     int              *vdwtype;
1103     real             *vdwparam;
1104     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1105     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1106     __m128           dummy_mask,cutoff_mask;
1107     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1108     __m128           one     = _mm_set1_ps(1.0);
1109     __m128           two     = _mm_set1_ps(2.0);
1110     x                = xx[0];
1111     f                = ff[0];
1112
1113     nri              = nlist->nri;
1114     iinr             = nlist->iinr;
1115     jindex           = nlist->jindex;
1116     jjnr             = nlist->jjnr;
1117     shiftidx         = nlist->shift;
1118     gid              = nlist->gid;
1119     shiftvec         = fr->shift_vec[0];
1120     fshift           = fr->fshift[0];
1121     facel            = _mm_set1_ps(fr->ic->epsfac);
1122     charge           = mdatoms->chargeA;
1123     nvdwtype         = fr->ntype;
1124     vdwparam         = fr->nbfp;
1125     vdwtype          = mdatoms->typeA;
1126
1127     /* Setup water-specific parameters */
1128     inr              = nlist->iinr[0];
1129     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1130     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1131     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1132     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1133
1134     jq1              = _mm_set1_ps(charge[inr+1]);
1135     jq2              = _mm_set1_ps(charge[inr+2]);
1136     jq3              = _mm_set1_ps(charge[inr+3]);
1137     vdwjidx0A        = 2*vdwtype[inr+0];
1138     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1139     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1140     qq11             = _mm_mul_ps(iq1,jq1);
1141     qq12             = _mm_mul_ps(iq1,jq2);
1142     qq13             = _mm_mul_ps(iq1,jq3);
1143     qq21             = _mm_mul_ps(iq2,jq1);
1144     qq22             = _mm_mul_ps(iq2,jq2);
1145     qq23             = _mm_mul_ps(iq2,jq3);
1146     qq31             = _mm_mul_ps(iq3,jq1);
1147     qq32             = _mm_mul_ps(iq3,jq2);
1148     qq33             = _mm_mul_ps(iq3,jq3);
1149
1150     /* Avoid stupid compiler warnings */
1151     jnrA = jnrB = jnrC = jnrD = 0;
1152     j_coord_offsetA = 0;
1153     j_coord_offsetB = 0;
1154     j_coord_offsetC = 0;
1155     j_coord_offsetD = 0;
1156
1157     outeriter        = 0;
1158     inneriter        = 0;
1159
1160     for(iidx=0;iidx<4*DIM;iidx++)
1161     {
1162         scratch[iidx] = 0.0;
1163     }
1164
1165     /* Start outer loop over neighborlists */
1166     for(iidx=0; iidx<nri; iidx++)
1167     {
1168         /* Load shift vector for this list */
1169         i_shift_offset   = DIM*shiftidx[iidx];
1170
1171         /* Load limits for loop over neighbors */
1172         j_index_start    = jindex[iidx];
1173         j_index_end      = jindex[iidx+1];
1174
1175         /* Get outer coordinate index */
1176         inr              = iinr[iidx];
1177         i_coord_offset   = DIM*inr;
1178
1179         /* Load i particle coords and add shift vector */
1180         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1181                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1182
1183         fix0             = _mm_setzero_ps();
1184         fiy0             = _mm_setzero_ps();
1185         fiz0             = _mm_setzero_ps();
1186         fix1             = _mm_setzero_ps();
1187         fiy1             = _mm_setzero_ps();
1188         fiz1             = _mm_setzero_ps();
1189         fix2             = _mm_setzero_ps();
1190         fiy2             = _mm_setzero_ps();
1191         fiz2             = _mm_setzero_ps();
1192         fix3             = _mm_setzero_ps();
1193         fiy3             = _mm_setzero_ps();
1194         fiz3             = _mm_setzero_ps();
1195
1196         /* Start inner kernel loop */
1197         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1198         {
1199
1200             /* Get j neighbor index, and coordinate index */
1201             jnrA             = jjnr[jidx];
1202             jnrB             = jjnr[jidx+1];
1203             jnrC             = jjnr[jidx+2];
1204             jnrD             = jjnr[jidx+3];
1205             j_coord_offsetA  = DIM*jnrA;
1206             j_coord_offsetB  = DIM*jnrB;
1207             j_coord_offsetC  = DIM*jnrC;
1208             j_coord_offsetD  = DIM*jnrD;
1209
1210             /* load j atom coordinates */
1211             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1212                                               x+j_coord_offsetC,x+j_coord_offsetD,
1213                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1214                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1215
1216             /* Calculate displacement vector */
1217             dx00             = _mm_sub_ps(ix0,jx0);
1218             dy00             = _mm_sub_ps(iy0,jy0);
1219             dz00             = _mm_sub_ps(iz0,jz0);
1220             dx11             = _mm_sub_ps(ix1,jx1);
1221             dy11             = _mm_sub_ps(iy1,jy1);
1222             dz11             = _mm_sub_ps(iz1,jz1);
1223             dx12             = _mm_sub_ps(ix1,jx2);
1224             dy12             = _mm_sub_ps(iy1,jy2);
1225             dz12             = _mm_sub_ps(iz1,jz2);
1226             dx13             = _mm_sub_ps(ix1,jx3);
1227             dy13             = _mm_sub_ps(iy1,jy3);
1228             dz13             = _mm_sub_ps(iz1,jz3);
1229             dx21             = _mm_sub_ps(ix2,jx1);
1230             dy21             = _mm_sub_ps(iy2,jy1);
1231             dz21             = _mm_sub_ps(iz2,jz1);
1232             dx22             = _mm_sub_ps(ix2,jx2);
1233             dy22             = _mm_sub_ps(iy2,jy2);
1234             dz22             = _mm_sub_ps(iz2,jz2);
1235             dx23             = _mm_sub_ps(ix2,jx3);
1236             dy23             = _mm_sub_ps(iy2,jy3);
1237             dz23             = _mm_sub_ps(iz2,jz3);
1238             dx31             = _mm_sub_ps(ix3,jx1);
1239             dy31             = _mm_sub_ps(iy3,jy1);
1240             dz31             = _mm_sub_ps(iz3,jz1);
1241             dx32             = _mm_sub_ps(ix3,jx2);
1242             dy32             = _mm_sub_ps(iy3,jy2);
1243             dz32             = _mm_sub_ps(iz3,jz2);
1244             dx33             = _mm_sub_ps(ix3,jx3);
1245             dy33             = _mm_sub_ps(iy3,jy3);
1246             dz33             = _mm_sub_ps(iz3,jz3);
1247
1248             /* Calculate squared distance and things based on it */
1249             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1250             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1251             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1252             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1253             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1254             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1255             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1256             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1257             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1258             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1259
1260             rinv11           = sse41_invsqrt_f(rsq11);
1261             rinv12           = sse41_invsqrt_f(rsq12);
1262             rinv13           = sse41_invsqrt_f(rsq13);
1263             rinv21           = sse41_invsqrt_f(rsq21);
1264             rinv22           = sse41_invsqrt_f(rsq22);
1265             rinv23           = sse41_invsqrt_f(rsq23);
1266             rinv31           = sse41_invsqrt_f(rsq31);
1267             rinv32           = sse41_invsqrt_f(rsq32);
1268             rinv33           = sse41_invsqrt_f(rsq33);
1269
1270             rinvsq00         = sse41_inv_f(rsq00);
1271             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1272             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1273             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1274             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1275             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1276             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1277             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1278             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1279             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1280
1281             fjx0             = _mm_setzero_ps();
1282             fjy0             = _mm_setzero_ps();
1283             fjz0             = _mm_setzero_ps();
1284             fjx1             = _mm_setzero_ps();
1285             fjy1             = _mm_setzero_ps();
1286             fjz1             = _mm_setzero_ps();
1287             fjx2             = _mm_setzero_ps();
1288             fjy2             = _mm_setzero_ps();
1289             fjz2             = _mm_setzero_ps();
1290             fjx3             = _mm_setzero_ps();
1291             fjy3             = _mm_setzero_ps();
1292             fjz3             = _mm_setzero_ps();
1293
1294             /**************************
1295              * CALCULATE INTERACTIONS *
1296              **************************/
1297
1298             /* LENNARD-JONES DISPERSION/REPULSION */
1299
1300             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1301             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1302
1303             fscal            = fvdw;
1304
1305             /* Calculate temporary vectorial force */
1306             tx               = _mm_mul_ps(fscal,dx00);
1307             ty               = _mm_mul_ps(fscal,dy00);
1308             tz               = _mm_mul_ps(fscal,dz00);
1309
1310             /* Update vectorial force */
1311             fix0             = _mm_add_ps(fix0,tx);
1312             fiy0             = _mm_add_ps(fiy0,ty);
1313             fiz0             = _mm_add_ps(fiz0,tz);
1314
1315             fjx0             = _mm_add_ps(fjx0,tx);
1316             fjy0             = _mm_add_ps(fjy0,ty);
1317             fjz0             = _mm_add_ps(fjz0,tz);
1318
1319             /**************************
1320              * CALCULATE INTERACTIONS *
1321              **************************/
1322
1323             /* COULOMB ELECTROSTATICS */
1324             velec            = _mm_mul_ps(qq11,rinv11);
1325             felec            = _mm_mul_ps(velec,rinvsq11);
1326
1327             fscal            = felec;
1328
1329             /* Calculate temporary vectorial force */
1330             tx               = _mm_mul_ps(fscal,dx11);
1331             ty               = _mm_mul_ps(fscal,dy11);
1332             tz               = _mm_mul_ps(fscal,dz11);
1333
1334             /* Update vectorial force */
1335             fix1             = _mm_add_ps(fix1,tx);
1336             fiy1             = _mm_add_ps(fiy1,ty);
1337             fiz1             = _mm_add_ps(fiz1,tz);
1338
1339             fjx1             = _mm_add_ps(fjx1,tx);
1340             fjy1             = _mm_add_ps(fjy1,ty);
1341             fjz1             = _mm_add_ps(fjz1,tz);
1342
1343             /**************************
1344              * CALCULATE INTERACTIONS *
1345              **************************/
1346
1347             /* COULOMB ELECTROSTATICS */
1348             velec            = _mm_mul_ps(qq12,rinv12);
1349             felec            = _mm_mul_ps(velec,rinvsq12);
1350
1351             fscal            = felec;
1352
1353             /* Calculate temporary vectorial force */
1354             tx               = _mm_mul_ps(fscal,dx12);
1355             ty               = _mm_mul_ps(fscal,dy12);
1356             tz               = _mm_mul_ps(fscal,dz12);
1357
1358             /* Update vectorial force */
1359             fix1             = _mm_add_ps(fix1,tx);
1360             fiy1             = _mm_add_ps(fiy1,ty);
1361             fiz1             = _mm_add_ps(fiz1,tz);
1362
1363             fjx2             = _mm_add_ps(fjx2,tx);
1364             fjy2             = _mm_add_ps(fjy2,ty);
1365             fjz2             = _mm_add_ps(fjz2,tz);
1366
1367             /**************************
1368              * CALCULATE INTERACTIONS *
1369              **************************/
1370
1371             /* COULOMB ELECTROSTATICS */
1372             velec            = _mm_mul_ps(qq13,rinv13);
1373             felec            = _mm_mul_ps(velec,rinvsq13);
1374
1375             fscal            = felec;
1376
1377             /* Calculate temporary vectorial force */
1378             tx               = _mm_mul_ps(fscal,dx13);
1379             ty               = _mm_mul_ps(fscal,dy13);
1380             tz               = _mm_mul_ps(fscal,dz13);
1381
1382             /* Update vectorial force */
1383             fix1             = _mm_add_ps(fix1,tx);
1384             fiy1             = _mm_add_ps(fiy1,ty);
1385             fiz1             = _mm_add_ps(fiz1,tz);
1386
1387             fjx3             = _mm_add_ps(fjx3,tx);
1388             fjy3             = _mm_add_ps(fjy3,ty);
1389             fjz3             = _mm_add_ps(fjz3,tz);
1390
1391             /**************************
1392              * CALCULATE INTERACTIONS *
1393              **************************/
1394
1395             /* COULOMB ELECTROSTATICS */
1396             velec            = _mm_mul_ps(qq21,rinv21);
1397             felec            = _mm_mul_ps(velec,rinvsq21);
1398
1399             fscal            = felec;
1400
1401             /* Calculate temporary vectorial force */
1402             tx               = _mm_mul_ps(fscal,dx21);
1403             ty               = _mm_mul_ps(fscal,dy21);
1404             tz               = _mm_mul_ps(fscal,dz21);
1405
1406             /* Update vectorial force */
1407             fix2             = _mm_add_ps(fix2,tx);
1408             fiy2             = _mm_add_ps(fiy2,ty);
1409             fiz2             = _mm_add_ps(fiz2,tz);
1410
1411             fjx1             = _mm_add_ps(fjx1,tx);
1412             fjy1             = _mm_add_ps(fjy1,ty);
1413             fjz1             = _mm_add_ps(fjz1,tz);
1414
1415             /**************************
1416              * CALCULATE INTERACTIONS *
1417              **************************/
1418
1419             /* COULOMB ELECTROSTATICS */
1420             velec            = _mm_mul_ps(qq22,rinv22);
1421             felec            = _mm_mul_ps(velec,rinvsq22);
1422
1423             fscal            = felec;
1424
1425             /* Calculate temporary vectorial force */
1426             tx               = _mm_mul_ps(fscal,dx22);
1427             ty               = _mm_mul_ps(fscal,dy22);
1428             tz               = _mm_mul_ps(fscal,dz22);
1429
1430             /* Update vectorial force */
1431             fix2             = _mm_add_ps(fix2,tx);
1432             fiy2             = _mm_add_ps(fiy2,ty);
1433             fiz2             = _mm_add_ps(fiz2,tz);
1434
1435             fjx2             = _mm_add_ps(fjx2,tx);
1436             fjy2             = _mm_add_ps(fjy2,ty);
1437             fjz2             = _mm_add_ps(fjz2,tz);
1438
1439             /**************************
1440              * CALCULATE INTERACTIONS *
1441              **************************/
1442
1443             /* COULOMB ELECTROSTATICS */
1444             velec            = _mm_mul_ps(qq23,rinv23);
1445             felec            = _mm_mul_ps(velec,rinvsq23);
1446
1447             fscal            = felec;
1448
1449             /* Calculate temporary vectorial force */
1450             tx               = _mm_mul_ps(fscal,dx23);
1451             ty               = _mm_mul_ps(fscal,dy23);
1452             tz               = _mm_mul_ps(fscal,dz23);
1453
1454             /* Update vectorial force */
1455             fix2             = _mm_add_ps(fix2,tx);
1456             fiy2             = _mm_add_ps(fiy2,ty);
1457             fiz2             = _mm_add_ps(fiz2,tz);
1458
1459             fjx3             = _mm_add_ps(fjx3,tx);
1460             fjy3             = _mm_add_ps(fjy3,ty);
1461             fjz3             = _mm_add_ps(fjz3,tz);
1462
1463             /**************************
1464              * CALCULATE INTERACTIONS *
1465              **************************/
1466
1467             /* COULOMB ELECTROSTATICS */
1468             velec            = _mm_mul_ps(qq31,rinv31);
1469             felec            = _mm_mul_ps(velec,rinvsq31);
1470
1471             fscal            = felec;
1472
1473             /* Calculate temporary vectorial force */
1474             tx               = _mm_mul_ps(fscal,dx31);
1475             ty               = _mm_mul_ps(fscal,dy31);
1476             tz               = _mm_mul_ps(fscal,dz31);
1477
1478             /* Update vectorial force */
1479             fix3             = _mm_add_ps(fix3,tx);
1480             fiy3             = _mm_add_ps(fiy3,ty);
1481             fiz3             = _mm_add_ps(fiz3,tz);
1482
1483             fjx1             = _mm_add_ps(fjx1,tx);
1484             fjy1             = _mm_add_ps(fjy1,ty);
1485             fjz1             = _mm_add_ps(fjz1,tz);
1486
1487             /**************************
1488              * CALCULATE INTERACTIONS *
1489              **************************/
1490
1491             /* COULOMB ELECTROSTATICS */
1492             velec            = _mm_mul_ps(qq32,rinv32);
1493             felec            = _mm_mul_ps(velec,rinvsq32);
1494
1495             fscal            = felec;
1496
1497             /* Calculate temporary vectorial force */
1498             tx               = _mm_mul_ps(fscal,dx32);
1499             ty               = _mm_mul_ps(fscal,dy32);
1500             tz               = _mm_mul_ps(fscal,dz32);
1501
1502             /* Update vectorial force */
1503             fix3             = _mm_add_ps(fix3,tx);
1504             fiy3             = _mm_add_ps(fiy3,ty);
1505             fiz3             = _mm_add_ps(fiz3,tz);
1506
1507             fjx2             = _mm_add_ps(fjx2,tx);
1508             fjy2             = _mm_add_ps(fjy2,ty);
1509             fjz2             = _mm_add_ps(fjz2,tz);
1510
1511             /**************************
1512              * CALCULATE INTERACTIONS *
1513              **************************/
1514
1515             /* COULOMB ELECTROSTATICS */
1516             velec            = _mm_mul_ps(qq33,rinv33);
1517             felec            = _mm_mul_ps(velec,rinvsq33);
1518
1519             fscal            = felec;
1520
1521             /* Calculate temporary vectorial force */
1522             tx               = _mm_mul_ps(fscal,dx33);
1523             ty               = _mm_mul_ps(fscal,dy33);
1524             tz               = _mm_mul_ps(fscal,dz33);
1525
1526             /* Update vectorial force */
1527             fix3             = _mm_add_ps(fix3,tx);
1528             fiy3             = _mm_add_ps(fiy3,ty);
1529             fiz3             = _mm_add_ps(fiz3,tz);
1530
1531             fjx3             = _mm_add_ps(fjx3,tx);
1532             fjy3             = _mm_add_ps(fjy3,ty);
1533             fjz3             = _mm_add_ps(fjz3,tz);
1534
1535             fjptrA             = f+j_coord_offsetA;
1536             fjptrB             = f+j_coord_offsetB;
1537             fjptrC             = f+j_coord_offsetC;
1538             fjptrD             = f+j_coord_offsetD;
1539
1540             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1541                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1542                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1543
1544             /* Inner loop uses 273 flops */
1545         }
1546
1547         if(jidx<j_index_end)
1548         {
1549
1550             /* Get j neighbor index, and coordinate index */
1551             jnrlistA         = jjnr[jidx];
1552             jnrlistB         = jjnr[jidx+1];
1553             jnrlistC         = jjnr[jidx+2];
1554             jnrlistD         = jjnr[jidx+3];
1555             /* Sign of each element will be negative for non-real atoms.
1556              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1557              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1558              */
1559             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1560             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1561             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1562             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1563             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1564             j_coord_offsetA  = DIM*jnrA;
1565             j_coord_offsetB  = DIM*jnrB;
1566             j_coord_offsetC  = DIM*jnrC;
1567             j_coord_offsetD  = DIM*jnrD;
1568
1569             /* load j atom coordinates */
1570             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1571                                               x+j_coord_offsetC,x+j_coord_offsetD,
1572                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1573                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1574
1575             /* Calculate displacement vector */
1576             dx00             = _mm_sub_ps(ix0,jx0);
1577             dy00             = _mm_sub_ps(iy0,jy0);
1578             dz00             = _mm_sub_ps(iz0,jz0);
1579             dx11             = _mm_sub_ps(ix1,jx1);
1580             dy11             = _mm_sub_ps(iy1,jy1);
1581             dz11             = _mm_sub_ps(iz1,jz1);
1582             dx12             = _mm_sub_ps(ix1,jx2);
1583             dy12             = _mm_sub_ps(iy1,jy2);
1584             dz12             = _mm_sub_ps(iz1,jz2);
1585             dx13             = _mm_sub_ps(ix1,jx3);
1586             dy13             = _mm_sub_ps(iy1,jy3);
1587             dz13             = _mm_sub_ps(iz1,jz3);
1588             dx21             = _mm_sub_ps(ix2,jx1);
1589             dy21             = _mm_sub_ps(iy2,jy1);
1590             dz21             = _mm_sub_ps(iz2,jz1);
1591             dx22             = _mm_sub_ps(ix2,jx2);
1592             dy22             = _mm_sub_ps(iy2,jy2);
1593             dz22             = _mm_sub_ps(iz2,jz2);
1594             dx23             = _mm_sub_ps(ix2,jx3);
1595             dy23             = _mm_sub_ps(iy2,jy3);
1596             dz23             = _mm_sub_ps(iz2,jz3);
1597             dx31             = _mm_sub_ps(ix3,jx1);
1598             dy31             = _mm_sub_ps(iy3,jy1);
1599             dz31             = _mm_sub_ps(iz3,jz1);
1600             dx32             = _mm_sub_ps(ix3,jx2);
1601             dy32             = _mm_sub_ps(iy3,jy2);
1602             dz32             = _mm_sub_ps(iz3,jz2);
1603             dx33             = _mm_sub_ps(ix3,jx3);
1604             dy33             = _mm_sub_ps(iy3,jy3);
1605             dz33             = _mm_sub_ps(iz3,jz3);
1606
1607             /* Calculate squared distance and things based on it */
1608             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1609             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1610             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1611             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1612             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1613             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1614             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1615             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1616             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1617             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1618
1619             rinv11           = sse41_invsqrt_f(rsq11);
1620             rinv12           = sse41_invsqrt_f(rsq12);
1621             rinv13           = sse41_invsqrt_f(rsq13);
1622             rinv21           = sse41_invsqrt_f(rsq21);
1623             rinv22           = sse41_invsqrt_f(rsq22);
1624             rinv23           = sse41_invsqrt_f(rsq23);
1625             rinv31           = sse41_invsqrt_f(rsq31);
1626             rinv32           = sse41_invsqrt_f(rsq32);
1627             rinv33           = sse41_invsqrt_f(rsq33);
1628
1629             rinvsq00         = sse41_inv_f(rsq00);
1630             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1631             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1632             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1633             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1634             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1635             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1636             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1637             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1638             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1639
1640             fjx0             = _mm_setzero_ps();
1641             fjy0             = _mm_setzero_ps();
1642             fjz0             = _mm_setzero_ps();
1643             fjx1             = _mm_setzero_ps();
1644             fjy1             = _mm_setzero_ps();
1645             fjz1             = _mm_setzero_ps();
1646             fjx2             = _mm_setzero_ps();
1647             fjy2             = _mm_setzero_ps();
1648             fjz2             = _mm_setzero_ps();
1649             fjx3             = _mm_setzero_ps();
1650             fjy3             = _mm_setzero_ps();
1651             fjz3             = _mm_setzero_ps();
1652
1653             /**************************
1654              * CALCULATE INTERACTIONS *
1655              **************************/
1656
1657             /* LENNARD-JONES DISPERSION/REPULSION */
1658
1659             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1660             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1661
1662             fscal            = fvdw;
1663
1664             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1665
1666             /* Calculate temporary vectorial force */
1667             tx               = _mm_mul_ps(fscal,dx00);
1668             ty               = _mm_mul_ps(fscal,dy00);
1669             tz               = _mm_mul_ps(fscal,dz00);
1670
1671             /* Update vectorial force */
1672             fix0             = _mm_add_ps(fix0,tx);
1673             fiy0             = _mm_add_ps(fiy0,ty);
1674             fiz0             = _mm_add_ps(fiz0,tz);
1675
1676             fjx0             = _mm_add_ps(fjx0,tx);
1677             fjy0             = _mm_add_ps(fjy0,ty);
1678             fjz0             = _mm_add_ps(fjz0,tz);
1679
1680             /**************************
1681              * CALCULATE INTERACTIONS *
1682              **************************/
1683
1684             /* COULOMB ELECTROSTATICS */
1685             velec            = _mm_mul_ps(qq11,rinv11);
1686             felec            = _mm_mul_ps(velec,rinvsq11);
1687
1688             fscal            = felec;
1689
1690             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1691
1692             /* Calculate temporary vectorial force */
1693             tx               = _mm_mul_ps(fscal,dx11);
1694             ty               = _mm_mul_ps(fscal,dy11);
1695             tz               = _mm_mul_ps(fscal,dz11);
1696
1697             /* Update vectorial force */
1698             fix1             = _mm_add_ps(fix1,tx);
1699             fiy1             = _mm_add_ps(fiy1,ty);
1700             fiz1             = _mm_add_ps(fiz1,tz);
1701
1702             fjx1             = _mm_add_ps(fjx1,tx);
1703             fjy1             = _mm_add_ps(fjy1,ty);
1704             fjz1             = _mm_add_ps(fjz1,tz);
1705
1706             /**************************
1707              * CALCULATE INTERACTIONS *
1708              **************************/
1709
1710             /* COULOMB ELECTROSTATICS */
1711             velec            = _mm_mul_ps(qq12,rinv12);
1712             felec            = _mm_mul_ps(velec,rinvsq12);
1713
1714             fscal            = felec;
1715
1716             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1717
1718             /* Calculate temporary vectorial force */
1719             tx               = _mm_mul_ps(fscal,dx12);
1720             ty               = _mm_mul_ps(fscal,dy12);
1721             tz               = _mm_mul_ps(fscal,dz12);
1722
1723             /* Update vectorial force */
1724             fix1             = _mm_add_ps(fix1,tx);
1725             fiy1             = _mm_add_ps(fiy1,ty);
1726             fiz1             = _mm_add_ps(fiz1,tz);
1727
1728             fjx2             = _mm_add_ps(fjx2,tx);
1729             fjy2             = _mm_add_ps(fjy2,ty);
1730             fjz2             = _mm_add_ps(fjz2,tz);
1731
1732             /**************************
1733              * CALCULATE INTERACTIONS *
1734              **************************/
1735
1736             /* COULOMB ELECTROSTATICS */
1737             velec            = _mm_mul_ps(qq13,rinv13);
1738             felec            = _mm_mul_ps(velec,rinvsq13);
1739
1740             fscal            = felec;
1741
1742             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1743
1744             /* Calculate temporary vectorial force */
1745             tx               = _mm_mul_ps(fscal,dx13);
1746             ty               = _mm_mul_ps(fscal,dy13);
1747             tz               = _mm_mul_ps(fscal,dz13);
1748
1749             /* Update vectorial force */
1750             fix1             = _mm_add_ps(fix1,tx);
1751             fiy1             = _mm_add_ps(fiy1,ty);
1752             fiz1             = _mm_add_ps(fiz1,tz);
1753
1754             fjx3             = _mm_add_ps(fjx3,tx);
1755             fjy3             = _mm_add_ps(fjy3,ty);
1756             fjz3             = _mm_add_ps(fjz3,tz);
1757
1758             /**************************
1759              * CALCULATE INTERACTIONS *
1760              **************************/
1761
1762             /* COULOMB ELECTROSTATICS */
1763             velec            = _mm_mul_ps(qq21,rinv21);
1764             felec            = _mm_mul_ps(velec,rinvsq21);
1765
1766             fscal            = felec;
1767
1768             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1769
1770             /* Calculate temporary vectorial force */
1771             tx               = _mm_mul_ps(fscal,dx21);
1772             ty               = _mm_mul_ps(fscal,dy21);
1773             tz               = _mm_mul_ps(fscal,dz21);
1774
1775             /* Update vectorial force */
1776             fix2             = _mm_add_ps(fix2,tx);
1777             fiy2             = _mm_add_ps(fiy2,ty);
1778             fiz2             = _mm_add_ps(fiz2,tz);
1779
1780             fjx1             = _mm_add_ps(fjx1,tx);
1781             fjy1             = _mm_add_ps(fjy1,ty);
1782             fjz1             = _mm_add_ps(fjz1,tz);
1783
1784             /**************************
1785              * CALCULATE INTERACTIONS *
1786              **************************/
1787
1788             /* COULOMB ELECTROSTATICS */
1789             velec            = _mm_mul_ps(qq22,rinv22);
1790             felec            = _mm_mul_ps(velec,rinvsq22);
1791
1792             fscal            = felec;
1793
1794             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1795
1796             /* Calculate temporary vectorial force */
1797             tx               = _mm_mul_ps(fscal,dx22);
1798             ty               = _mm_mul_ps(fscal,dy22);
1799             tz               = _mm_mul_ps(fscal,dz22);
1800
1801             /* Update vectorial force */
1802             fix2             = _mm_add_ps(fix2,tx);
1803             fiy2             = _mm_add_ps(fiy2,ty);
1804             fiz2             = _mm_add_ps(fiz2,tz);
1805
1806             fjx2             = _mm_add_ps(fjx2,tx);
1807             fjy2             = _mm_add_ps(fjy2,ty);
1808             fjz2             = _mm_add_ps(fjz2,tz);
1809
1810             /**************************
1811              * CALCULATE INTERACTIONS *
1812              **************************/
1813
1814             /* COULOMB ELECTROSTATICS */
1815             velec            = _mm_mul_ps(qq23,rinv23);
1816             felec            = _mm_mul_ps(velec,rinvsq23);
1817
1818             fscal            = felec;
1819
1820             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1821
1822             /* Calculate temporary vectorial force */
1823             tx               = _mm_mul_ps(fscal,dx23);
1824             ty               = _mm_mul_ps(fscal,dy23);
1825             tz               = _mm_mul_ps(fscal,dz23);
1826
1827             /* Update vectorial force */
1828             fix2             = _mm_add_ps(fix2,tx);
1829             fiy2             = _mm_add_ps(fiy2,ty);
1830             fiz2             = _mm_add_ps(fiz2,tz);
1831
1832             fjx3             = _mm_add_ps(fjx3,tx);
1833             fjy3             = _mm_add_ps(fjy3,ty);
1834             fjz3             = _mm_add_ps(fjz3,tz);
1835
1836             /**************************
1837              * CALCULATE INTERACTIONS *
1838              **************************/
1839
1840             /* COULOMB ELECTROSTATICS */
1841             velec            = _mm_mul_ps(qq31,rinv31);
1842             felec            = _mm_mul_ps(velec,rinvsq31);
1843
1844             fscal            = felec;
1845
1846             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1847
1848             /* Calculate temporary vectorial force */
1849             tx               = _mm_mul_ps(fscal,dx31);
1850             ty               = _mm_mul_ps(fscal,dy31);
1851             tz               = _mm_mul_ps(fscal,dz31);
1852
1853             /* Update vectorial force */
1854             fix3             = _mm_add_ps(fix3,tx);
1855             fiy3             = _mm_add_ps(fiy3,ty);
1856             fiz3             = _mm_add_ps(fiz3,tz);
1857
1858             fjx1             = _mm_add_ps(fjx1,tx);
1859             fjy1             = _mm_add_ps(fjy1,ty);
1860             fjz1             = _mm_add_ps(fjz1,tz);
1861
1862             /**************************
1863              * CALCULATE INTERACTIONS *
1864              **************************/
1865
1866             /* COULOMB ELECTROSTATICS */
1867             velec            = _mm_mul_ps(qq32,rinv32);
1868             felec            = _mm_mul_ps(velec,rinvsq32);
1869
1870             fscal            = felec;
1871
1872             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1873
1874             /* Calculate temporary vectorial force */
1875             tx               = _mm_mul_ps(fscal,dx32);
1876             ty               = _mm_mul_ps(fscal,dy32);
1877             tz               = _mm_mul_ps(fscal,dz32);
1878
1879             /* Update vectorial force */
1880             fix3             = _mm_add_ps(fix3,tx);
1881             fiy3             = _mm_add_ps(fiy3,ty);
1882             fiz3             = _mm_add_ps(fiz3,tz);
1883
1884             fjx2             = _mm_add_ps(fjx2,tx);
1885             fjy2             = _mm_add_ps(fjy2,ty);
1886             fjz2             = _mm_add_ps(fjz2,tz);
1887
1888             /**************************
1889              * CALCULATE INTERACTIONS *
1890              **************************/
1891
1892             /* COULOMB ELECTROSTATICS */
1893             velec            = _mm_mul_ps(qq33,rinv33);
1894             felec            = _mm_mul_ps(velec,rinvsq33);
1895
1896             fscal            = felec;
1897
1898             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1899
1900             /* Calculate temporary vectorial force */
1901             tx               = _mm_mul_ps(fscal,dx33);
1902             ty               = _mm_mul_ps(fscal,dy33);
1903             tz               = _mm_mul_ps(fscal,dz33);
1904
1905             /* Update vectorial force */
1906             fix3             = _mm_add_ps(fix3,tx);
1907             fiy3             = _mm_add_ps(fiy3,ty);
1908             fiz3             = _mm_add_ps(fiz3,tz);
1909
1910             fjx3             = _mm_add_ps(fjx3,tx);
1911             fjy3             = _mm_add_ps(fjy3,ty);
1912             fjz3             = _mm_add_ps(fjz3,tz);
1913
1914             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1915             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1916             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1917             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1918
1919             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1920                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1921                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1922
1923             /* Inner loop uses 273 flops */
1924         }
1925
1926         /* End of innermost loop */
1927
1928         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1929                                               f+i_coord_offset,fshift+i_shift_offset);
1930
1931         /* Increment number of inner iterations */
1932         inneriter                  += j_index_end - j_index_start;
1933
1934         /* Outer loop uses 24 flops */
1935     }
1936
1937     /* Increment number of outer iterations */
1938     outeriter        += nri;
1939
1940     /* Update outer/inner flops */
1941
1942     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*273);
1943 }