src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_ElecCoul_VdwLJ_GeomW3W3_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 "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "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_GeomW3W3_VF_sse4_1_single
  52  * Electrostatics interaction: Coulomb
  53  * VdW interaction:            LennardJones
  54  * Geometry:                   Water3-Water3
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecCoul_VdwLJ_GeomW3W3_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              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  90     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  91     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  92     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  93     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  94     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  95     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  96     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  97     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  98     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  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           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 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           velec,felec,velecsum,facel,crf,krf,krf2;
 105     real             *charge;
 106     int              nvdwtype;
 107     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 108     int              *vdwtype;
 109     real             *vdwparam;
 110     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 111     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 112     __m128           dummy_mask,cutoff_mask;
 113     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 114     __m128           one     = _mm_set1_ps(1.0);
 115     __m128           two     = _mm_set1_ps(2.0);
 116     x                = xx[0];
 117     f                = ff[0];
 118
 119     nri              = nlist->nri;
 120     iinr             = nlist->iinr;
 121     jindex           = nlist->jindex;
 122     jjnr             = nlist->jjnr;
 123     shiftidx         = nlist->shift;
 124     gid              = nlist->gid;
 125     shiftvec         = fr->shift_vec[0];
 126     fshift           = fr->fshift[0];
 127     facel            = _mm_set1_ps(fr->epsfac);
 128     charge           = mdatoms->chargeA;
 129     nvdwtype         = fr->ntype;
 130     vdwparam         = fr->nbfp;
 131     vdwtype          = mdatoms->typeA;
 132
 133     /* Setup water-specific parameters */
 134     inr              = nlist->iinr[0];
 135     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 136     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 137     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 138     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 139
 140     jq0              = _mm_set1_ps(charge[inr+0]);
 141     jq1              = _mm_set1_ps(charge[inr+1]);
 142     jq2              = _mm_set1_ps(charge[inr+2]);
 143     vdwjidx0A        = 2*vdwtype[inr+0];
 144     qq00             = _mm_mul_ps(iq0,jq0);
 145     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 146     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 147     qq01             = _mm_mul_ps(iq0,jq1);
 148     qq02             = _mm_mul_ps(iq0,jq2);
 149     qq10             = _mm_mul_ps(iq1,jq0);
 150     qq11             = _mm_mul_ps(iq1,jq1);
 151     qq12             = _mm_mul_ps(iq1,jq2);
 152     qq20             = _mm_mul_ps(iq2,jq0);
 153     qq21             = _mm_mul_ps(iq2,jq1);
 154     qq22             = _mm_mul_ps(iq2,jq2);
 155
 156     /* Avoid stupid compiler warnings */
 157     jnrA = jnrB = jnrC = jnrD = 0;
 158     j_coord_offsetA = 0;
 159     j_coord_offsetB = 0;
 160     j_coord_offsetC = 0;
 161     j_coord_offsetD = 0;
 162
 163     outeriter        = 0;
 164     inneriter        = 0;
 165
 166     for(iidx=0;iidx<4*DIM;iidx++)
 167     {
 168         scratch[iidx] = 0.0;
 169     }
 170
 171     /* Start outer loop over neighborlists */
 172     for(iidx=0; iidx<nri; iidx++)
 173     {
 174         /* Load shift vector for this list */
 175         i_shift_offset   = DIM*shiftidx[iidx];
 176
 177         /* Load limits for loop over neighbors */
 178         j_index_start    = jindex[iidx];
 179         j_index_end      = jindex[iidx+1];
 180
 181         /* Get outer coordinate index */
 182         inr              = iinr[iidx];
 183         i_coord_offset   = DIM*inr;
 184
 185         /* Load i particle coords and add shift vector */
 186         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 187                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 188
 189         fix0             = _mm_setzero_ps();
 190         fiy0             = _mm_setzero_ps();
 191         fiz0             = _mm_setzero_ps();
 192         fix1             = _mm_setzero_ps();
 193         fiy1             = _mm_setzero_ps();
 194         fiz1             = _mm_setzero_ps();
 195         fix2             = _mm_setzero_ps();
 196         fiy2             = _mm_setzero_ps();
 197         fiz2             = _mm_setzero_ps();
 198
 199         /* Reset potential sums */
 200         velecsum         = _mm_setzero_ps();
 201         vvdwsum          = _mm_setzero_ps();
 202
 203         /* Start inner kernel loop */
 204         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 205         {
 206
 207             /* Get j neighbor index, and coordinate index */
 208             jnrA             = jjnr[jidx];
 209             jnrB             = jjnr[jidx+1];
 210             jnrC             = jjnr[jidx+2];
 211             jnrD             = jjnr[jidx+3];
 212             j_coord_offsetA  = DIM*jnrA;
 213             j_coord_offsetB  = DIM*jnrB;
 214             j_coord_offsetC  = DIM*jnrC;
 215             j_coord_offsetD  = DIM*jnrD;
 216
 217             /* load j atom coordinates */
 218             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 219                                               x+j_coord_offsetC,x+j_coord_offsetD,
 220                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 221
 222             /* Calculate displacement vector */
 223             dx00             = _mm_sub_ps(ix0,jx0);
 224             dy00             = _mm_sub_ps(iy0,jy0);
 225             dz00             = _mm_sub_ps(iz0,jz0);
 226             dx01             = _mm_sub_ps(ix0,jx1);
 227             dy01             = _mm_sub_ps(iy0,jy1);
 228             dz01             = _mm_sub_ps(iz0,jz1);
 229             dx02             = _mm_sub_ps(ix0,jx2);
 230             dy02             = _mm_sub_ps(iy0,jy2);
 231             dz02             = _mm_sub_ps(iz0,jz2);
 232             dx10             = _mm_sub_ps(ix1,jx0);
 233             dy10             = _mm_sub_ps(iy1,jy0);
 234             dz10             = _mm_sub_ps(iz1,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             dx20             = _mm_sub_ps(ix2,jx0);
 242             dy20             = _mm_sub_ps(iy2,jy0);
 243             dz20             = _mm_sub_ps(iz2,jz0);
 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
 251             /* Calculate squared distance and things based on it */
 252             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 253             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 254             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 255             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 256             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 257             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 258             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 259             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 260             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 261
 262             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 263             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 264             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 265             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 266             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 267             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 268             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 269             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 270             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 271
 272             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 273             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 274             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 275             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 276             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 277             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 278             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 279             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 280             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 281
 282             fjx0             = _mm_setzero_ps();
 283             fjy0             = _mm_setzero_ps();
 284             fjz0             = _mm_setzero_ps();
 285             fjx1             = _mm_setzero_ps();
 286             fjy1             = _mm_setzero_ps();
 287             fjz1             = _mm_setzero_ps();
 288             fjx2             = _mm_setzero_ps();
 289             fjy2             = _mm_setzero_ps();
 290             fjz2             = _mm_setzero_ps();
 291
 292             /**************************
 293              * CALCULATE INTERACTIONS *
 294              **************************/
 295
 296             /* COULOMB ELECTROSTATICS */
 297             velec            = _mm_mul_ps(qq00,rinv00);
 298             felec            = _mm_mul_ps(velec,rinvsq00);
 299
 300             /* LENNARD-JONES DISPERSION/REPULSION */
 301
 302             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 303             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 304             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 305             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 306             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 307
 308             /* Update potential sum for this i atom from the interaction with this j atom. */
 309             velecsum         = _mm_add_ps(velecsum,velec);
 310             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 311
 312             fscal            = _mm_add_ps(felec,fvdw);
 313
 314             /* Calculate temporary vectorial force */
 315             tx               = _mm_mul_ps(fscal,dx00);
 316             ty               = _mm_mul_ps(fscal,dy00);
 317             tz               = _mm_mul_ps(fscal,dz00);
 318
 319             /* Update vectorial force */
 320             fix0             = _mm_add_ps(fix0,tx);
 321             fiy0             = _mm_add_ps(fiy0,ty);
 322             fiz0             = _mm_add_ps(fiz0,tz);
 323
 324             fjx0             = _mm_add_ps(fjx0,tx);
 325             fjy0             = _mm_add_ps(fjy0,ty);
 326             fjz0             = _mm_add_ps(fjz0,tz);
 327
 328             /**************************
 329              * CALCULATE INTERACTIONS *
 330              **************************/
 331
 332             /* COULOMB ELECTROSTATICS */
 333             velec            = _mm_mul_ps(qq01,rinv01);
 334             felec            = _mm_mul_ps(velec,rinvsq01);
 335
 336             /* Update potential sum for this i atom from the interaction with this j atom. */
 337             velecsum         = _mm_add_ps(velecsum,velec);
 338
 339             fscal            = felec;
 340
 341             /* Calculate temporary vectorial force */
 342             tx               = _mm_mul_ps(fscal,dx01);
 343             ty               = _mm_mul_ps(fscal,dy01);
 344             tz               = _mm_mul_ps(fscal,dz01);
 345
 346             /* Update vectorial force */
 347             fix0             = _mm_add_ps(fix0,tx);
 348             fiy0             = _mm_add_ps(fiy0,ty);
 349             fiz0             = _mm_add_ps(fiz0,tz);
 350
 351             fjx1             = _mm_add_ps(fjx1,tx);
 352             fjy1             = _mm_add_ps(fjy1,ty);
 353             fjz1             = _mm_add_ps(fjz1,tz);
 354
 355             /**************************
 356              * CALCULATE INTERACTIONS *
 357              **************************/
 358
 359             /* COULOMB ELECTROSTATICS */
 360             velec            = _mm_mul_ps(qq02,rinv02);
 361             felec            = _mm_mul_ps(velec,rinvsq02);
 362
 363             /* Update potential sum for this i atom from the interaction with this j atom. */
 364             velecsum         = _mm_add_ps(velecsum,velec);
 365
 366             fscal            = felec;
 367
 368             /* Calculate temporary vectorial force */
 369             tx               = _mm_mul_ps(fscal,dx02);
 370             ty               = _mm_mul_ps(fscal,dy02);
 371             tz               = _mm_mul_ps(fscal,dz02);
 372
 373             /* Update vectorial force */
 374             fix0             = _mm_add_ps(fix0,tx);
 375             fiy0             = _mm_add_ps(fiy0,ty);
 376             fiz0             = _mm_add_ps(fiz0,tz);
 377
 378             fjx2             = _mm_add_ps(fjx2,tx);
 379             fjy2             = _mm_add_ps(fjy2,ty);
 380             fjz2             = _mm_add_ps(fjz2,tz);
 381
 382             /**************************
 383              * CALCULATE INTERACTIONS *
 384              **************************/
 385
 386             /* COULOMB ELECTROSTATICS */
 387             velec            = _mm_mul_ps(qq10,rinv10);
 388             felec            = _mm_mul_ps(velec,rinvsq10);
 389
 390             /* Update potential sum for this i atom from the interaction with this j atom. */
 391             velecsum         = _mm_add_ps(velecsum,velec);
 392
 393             fscal            = felec;
 394
 395             /* Calculate temporary vectorial force */
 396             tx               = _mm_mul_ps(fscal,dx10);
 397             ty               = _mm_mul_ps(fscal,dy10);
 398             tz               = _mm_mul_ps(fscal,dz10);
 399
 400             /* Update vectorial force */
 401             fix1             = _mm_add_ps(fix1,tx);
 402             fiy1             = _mm_add_ps(fiy1,ty);
 403             fiz1             = _mm_add_ps(fiz1,tz);
 404
 405             fjx0             = _mm_add_ps(fjx0,tx);
 406             fjy0             = _mm_add_ps(fjy0,ty);
 407             fjz0             = _mm_add_ps(fjz0,tz);
 408
 409             /**************************
 410              * CALCULATE INTERACTIONS *
 411              **************************/
 412
 413             /* COULOMB ELECTROSTATICS */
 414             velec            = _mm_mul_ps(qq11,rinv11);
 415             felec            = _mm_mul_ps(velec,rinvsq11);
 416
 417             /* Update potential sum for this i atom from the interaction with this j atom. */
 418             velecsum         = _mm_add_ps(velecsum,velec);
 419
 420             fscal            = felec;
 421
 422             /* Calculate temporary vectorial force */
 423             tx               = _mm_mul_ps(fscal,dx11);
 424             ty               = _mm_mul_ps(fscal,dy11);
 425             tz               = _mm_mul_ps(fscal,dz11);
 426
 427             /* Update vectorial force */
 428             fix1             = _mm_add_ps(fix1,tx);
 429             fiy1             = _mm_add_ps(fiy1,ty);
 430             fiz1             = _mm_add_ps(fiz1,tz);
 431
 432             fjx1             = _mm_add_ps(fjx1,tx);
 433             fjy1             = _mm_add_ps(fjy1,ty);
 434             fjz1             = _mm_add_ps(fjz1,tz);
 435
 436             /**************************
 437              * CALCULATE INTERACTIONS *
 438              **************************/
 439
 440             /* COULOMB ELECTROSTATICS */
 441             velec            = _mm_mul_ps(qq12,rinv12);
 442             felec            = _mm_mul_ps(velec,rinvsq12);
 443
 444             /* Update potential sum for this i atom from the interaction with this j atom. */
 445             velecsum         = _mm_add_ps(velecsum,velec);
 446
 447             fscal            = felec;
 448
 449             /* Calculate temporary vectorial force */
 450             tx               = _mm_mul_ps(fscal,dx12);
 451             ty               = _mm_mul_ps(fscal,dy12);
 452             tz               = _mm_mul_ps(fscal,dz12);
 453
 454             /* Update vectorial force */
 455             fix1             = _mm_add_ps(fix1,tx);
 456             fiy1             = _mm_add_ps(fiy1,ty);
 457             fiz1             = _mm_add_ps(fiz1,tz);
 458
 459             fjx2             = _mm_add_ps(fjx2,tx);
 460             fjy2             = _mm_add_ps(fjy2,ty);
 461             fjz2             = _mm_add_ps(fjz2,tz);
 462
 463             /**************************
 464              * CALCULATE INTERACTIONS *
 465              **************************/
 466
 467             /* COULOMB ELECTROSTATICS */
 468             velec            = _mm_mul_ps(qq20,rinv20);
 469             felec            = _mm_mul_ps(velec,rinvsq20);
 470
 471             /* Update potential sum for this i atom from the interaction with this j atom. */
 472             velecsum         = _mm_add_ps(velecsum,velec);
 473
 474             fscal            = felec;
 475
 476             /* Calculate temporary vectorial force */
 477             tx               = _mm_mul_ps(fscal,dx20);
 478             ty               = _mm_mul_ps(fscal,dy20);
 479             tz               = _mm_mul_ps(fscal,dz20);
 480
 481             /* Update vectorial force */
 482             fix2             = _mm_add_ps(fix2,tx);
 483             fiy2             = _mm_add_ps(fiy2,ty);
 484             fiz2             = _mm_add_ps(fiz2,tz);
 485
 486             fjx0             = _mm_add_ps(fjx0,tx);
 487             fjy0             = _mm_add_ps(fjy0,ty);
 488             fjz0             = _mm_add_ps(fjz0,tz);
 489
 490             /**************************
 491              * CALCULATE INTERACTIONS *
 492              **************************/
 493
 494             /* COULOMB ELECTROSTATICS */
 495             velec            = _mm_mul_ps(qq21,rinv21);
 496             felec            = _mm_mul_ps(velec,rinvsq21);
 497
 498             /* Update potential sum for this i atom from the interaction with this j atom. */
 499             velecsum         = _mm_add_ps(velecsum,velec);
 500
 501             fscal            = felec;
 502
 503             /* Calculate temporary vectorial force */
 504             tx               = _mm_mul_ps(fscal,dx21);
 505             ty               = _mm_mul_ps(fscal,dy21);
 506             tz               = _mm_mul_ps(fscal,dz21);
 507
 508             /* Update vectorial force */
 509             fix2             = _mm_add_ps(fix2,tx);
 510             fiy2             = _mm_add_ps(fiy2,ty);
 511             fiz2             = _mm_add_ps(fiz2,tz);
 512
 513             fjx1             = _mm_add_ps(fjx1,tx);
 514             fjy1             = _mm_add_ps(fjy1,ty);
 515             fjz1             = _mm_add_ps(fjz1,tz);
 516
 517             /**************************
 518              * CALCULATE INTERACTIONS *
 519              **************************/
 520
 521             /* COULOMB ELECTROSTATICS */
 522             velec            = _mm_mul_ps(qq22,rinv22);
 523             felec            = _mm_mul_ps(velec,rinvsq22);
 524
 525             /* Update potential sum for this i atom from the interaction with this j atom. */
 526             velecsum         = _mm_add_ps(velecsum,velec);
 527
 528             fscal            = felec;
 529
 530             /* Calculate temporary vectorial force */
 531             tx               = _mm_mul_ps(fscal,dx22);
 532             ty               = _mm_mul_ps(fscal,dy22);
 533             tz               = _mm_mul_ps(fscal,dz22);
 534
 535             /* Update vectorial force */
 536             fix2             = _mm_add_ps(fix2,tx);
 537             fiy2             = _mm_add_ps(fiy2,ty);
 538             fiz2             = _mm_add_ps(fiz2,tz);
 539
 540             fjx2             = _mm_add_ps(fjx2,tx);
 541             fjy2             = _mm_add_ps(fjy2,ty);
 542             fjz2             = _mm_add_ps(fjz2,tz);
 543
 544             fjptrA             = f+j_coord_offsetA;
 545             fjptrB             = f+j_coord_offsetB;
 546             fjptrC             = f+j_coord_offsetC;
 547             fjptrD             = f+j_coord_offsetD;
 548
 549             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 550                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 551
 552             /* Inner loop uses 264 flops */
 553         }
 554
 555         if(jidx<j_index_end)
 556         {
 557
 558             /* Get j neighbor index, and coordinate index */
 559             jnrlistA         = jjnr[jidx];
 560             jnrlistB         = jjnr[jidx+1];
 561             jnrlistC         = jjnr[jidx+2];
 562             jnrlistD         = jjnr[jidx+3];
 563             /* Sign of each element will be negative for non-real atoms.
 564              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 565              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 566              */
 567             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 568             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 569             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 570             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 571             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 572             j_coord_offsetA  = DIM*jnrA;
 573             j_coord_offsetB  = DIM*jnrB;
 574             j_coord_offsetC  = DIM*jnrC;
 575             j_coord_offsetD  = DIM*jnrD;
 576
 577             /* load j atom coordinates */
 578             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 579                                               x+j_coord_offsetC,x+j_coord_offsetD,
 580                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 581
 582             /* Calculate displacement vector */
 583             dx00             = _mm_sub_ps(ix0,jx0);
 584             dy00             = _mm_sub_ps(iy0,jy0);
 585             dz00             = _mm_sub_ps(iz0,jz0);
 586             dx01             = _mm_sub_ps(ix0,jx1);
 587             dy01             = _mm_sub_ps(iy0,jy1);
 588             dz01             = _mm_sub_ps(iz0,jz1);
 589             dx02             = _mm_sub_ps(ix0,jx2);
 590             dy02             = _mm_sub_ps(iy0,jy2);
 591             dz02             = _mm_sub_ps(iz0,jz2);
 592             dx10             = _mm_sub_ps(ix1,jx0);
 593             dy10             = _mm_sub_ps(iy1,jy0);
 594             dz10             = _mm_sub_ps(iz1,jz0);
 595             dx11             = _mm_sub_ps(ix1,jx1);
 596             dy11             = _mm_sub_ps(iy1,jy1);
 597             dz11             = _mm_sub_ps(iz1,jz1);
 598             dx12             = _mm_sub_ps(ix1,jx2);
 599             dy12             = _mm_sub_ps(iy1,jy2);
 600             dz12             = _mm_sub_ps(iz1,jz2);
 601             dx20             = _mm_sub_ps(ix2,jx0);
 602             dy20             = _mm_sub_ps(iy2,jy0);
 603             dz20             = _mm_sub_ps(iz2,jz0);
 604             dx21             = _mm_sub_ps(ix2,jx1);
 605             dy21             = _mm_sub_ps(iy2,jy1);
 606             dz21             = _mm_sub_ps(iz2,jz1);
 607             dx22             = _mm_sub_ps(ix2,jx2);
 608             dy22             = _mm_sub_ps(iy2,jy2);
 609             dz22             = _mm_sub_ps(iz2,jz2);
 610
 611             /* Calculate squared distance and things based on it */
 612             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 613             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 614             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 615             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 616             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 617             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 618             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 619             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 620             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 621
 622             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 623             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 624             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 625             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 626             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 627             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 628             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 629             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 630             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 631
 632             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 633             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 634             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 635             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 636             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 637             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 638             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 639             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 640             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 641
 642             fjx0             = _mm_setzero_ps();
 643             fjy0             = _mm_setzero_ps();
 644             fjz0             = _mm_setzero_ps();
 645             fjx1             = _mm_setzero_ps();
 646             fjy1             = _mm_setzero_ps();
 647             fjz1             = _mm_setzero_ps();
 648             fjx2             = _mm_setzero_ps();
 649             fjy2             = _mm_setzero_ps();
 650             fjz2             = _mm_setzero_ps();
 651
 652             /**************************
 653              * CALCULATE INTERACTIONS *
 654              **************************/
 655
 656             /* COULOMB ELECTROSTATICS */
 657             velec            = _mm_mul_ps(qq00,rinv00);
 658             felec            = _mm_mul_ps(velec,rinvsq00);
 659
 660             /* LENNARD-JONES DISPERSION/REPULSION */
 661
 662             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 663             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 664             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 665             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 666             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 667
 668             /* Update potential sum for this i atom from the interaction with this j atom. */
 669             velec            = _mm_andnot_ps(dummy_mask,velec);
 670             velecsum         = _mm_add_ps(velecsum,velec);
 671             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 672             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 673
 674             fscal            = _mm_add_ps(felec,fvdw);
 675
 676             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 677
 678             /* Calculate temporary vectorial force */
 679             tx               = _mm_mul_ps(fscal,dx00);
 680             ty               = _mm_mul_ps(fscal,dy00);
 681             tz               = _mm_mul_ps(fscal,dz00);
 682
 683             /* Update vectorial force */
 684             fix0             = _mm_add_ps(fix0,tx);
 685             fiy0             = _mm_add_ps(fiy0,ty);
 686             fiz0             = _mm_add_ps(fiz0,tz);
 687
 688             fjx0             = _mm_add_ps(fjx0,tx);
 689             fjy0             = _mm_add_ps(fjy0,ty);
 690             fjz0             = _mm_add_ps(fjz0,tz);
 691
 692             /**************************
 693              * CALCULATE INTERACTIONS *
 694              **************************/
 695
 696             /* COULOMB ELECTROSTATICS */
 697             velec            = _mm_mul_ps(qq01,rinv01);
 698             felec            = _mm_mul_ps(velec,rinvsq01);
 699
 700             /* Update potential sum for this i atom from the interaction with this j atom. */
 701             velec            = _mm_andnot_ps(dummy_mask,velec);
 702             velecsum         = _mm_add_ps(velecsum,velec);
 703
 704             fscal            = felec;
 705
 706             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 707
 708             /* Calculate temporary vectorial force */
 709             tx               = _mm_mul_ps(fscal,dx01);
 710             ty               = _mm_mul_ps(fscal,dy01);
 711             tz               = _mm_mul_ps(fscal,dz01);
 712
 713             /* Update vectorial force */
 714             fix0             = _mm_add_ps(fix0,tx);
 715             fiy0             = _mm_add_ps(fiy0,ty);
 716             fiz0             = _mm_add_ps(fiz0,tz);
 717
 718             fjx1             = _mm_add_ps(fjx1,tx);
 719             fjy1             = _mm_add_ps(fjy1,ty);
 720             fjz1             = _mm_add_ps(fjz1,tz);
 721
 722             /**************************
 723              * CALCULATE INTERACTIONS *
 724              **************************/
 725
 726             /* COULOMB ELECTROSTATICS */
 727             velec            = _mm_mul_ps(qq02,rinv02);
 728             felec            = _mm_mul_ps(velec,rinvsq02);
 729
 730             /* Update potential sum for this i atom from the interaction with this j atom. */
 731             velec            = _mm_andnot_ps(dummy_mask,velec);
 732             velecsum         = _mm_add_ps(velecsum,velec);
 733
 734             fscal            = felec;
 735
 736             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 737
 738             /* Calculate temporary vectorial force */
 739             tx               = _mm_mul_ps(fscal,dx02);
 740             ty               = _mm_mul_ps(fscal,dy02);
 741             tz               = _mm_mul_ps(fscal,dz02);
 742
 743             /* Update vectorial force */
 744             fix0             = _mm_add_ps(fix0,tx);
 745             fiy0             = _mm_add_ps(fiy0,ty);
 746             fiz0             = _mm_add_ps(fiz0,tz);
 747
 748             fjx2             = _mm_add_ps(fjx2,tx);
 749             fjy2             = _mm_add_ps(fjy2,ty);
 750             fjz2             = _mm_add_ps(fjz2,tz);
 751
 752             /**************************
 753              * CALCULATE INTERACTIONS *
 754              **************************/
 755
 756             /* COULOMB ELECTROSTATICS */
 757             velec            = _mm_mul_ps(qq10,rinv10);
 758             felec            = _mm_mul_ps(velec,rinvsq10);
 759
 760             /* Update potential sum for this i atom from the interaction with this j atom. */
 761             velec            = _mm_andnot_ps(dummy_mask,velec);
 762             velecsum         = _mm_add_ps(velecsum,velec);
 763
 764             fscal            = felec;
 765
 766             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 767
 768             /* Calculate temporary vectorial force */
 769             tx               = _mm_mul_ps(fscal,dx10);
 770             ty               = _mm_mul_ps(fscal,dy10);
 771             tz               = _mm_mul_ps(fscal,dz10);
 772
 773             /* Update vectorial force */
 774             fix1             = _mm_add_ps(fix1,tx);
 775             fiy1             = _mm_add_ps(fiy1,ty);
 776             fiz1             = _mm_add_ps(fiz1,tz);
 777
 778             fjx0             = _mm_add_ps(fjx0,tx);
 779             fjy0             = _mm_add_ps(fjy0,ty);
 780             fjz0             = _mm_add_ps(fjz0,tz);
 781
 782             /**************************
 783              * CALCULATE INTERACTIONS *
 784              **************************/
 785
 786             /* COULOMB ELECTROSTATICS */
 787             velec            = _mm_mul_ps(qq11,rinv11);
 788             felec            = _mm_mul_ps(velec,rinvsq11);
 789
 790             /* Update potential sum for this i atom from the interaction with this j atom. */
 791             velec            = _mm_andnot_ps(dummy_mask,velec);
 792             velecsum         = _mm_add_ps(velecsum,velec);
 793
 794             fscal            = felec;
 795
 796             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 797
 798             /* Calculate temporary vectorial force */
 799             tx               = _mm_mul_ps(fscal,dx11);
 800             ty               = _mm_mul_ps(fscal,dy11);
 801             tz               = _mm_mul_ps(fscal,dz11);
 802
 803             /* Update vectorial force */
 804             fix1             = _mm_add_ps(fix1,tx);
 805             fiy1             = _mm_add_ps(fiy1,ty);
 806             fiz1             = _mm_add_ps(fiz1,tz);
 807
 808             fjx1             = _mm_add_ps(fjx1,tx);
 809             fjy1             = _mm_add_ps(fjy1,ty);
 810             fjz1             = _mm_add_ps(fjz1,tz);
 811
 812             /**************************
 813              * CALCULATE INTERACTIONS *
 814              **************************/
 815
 816             /* COULOMB ELECTROSTATICS */
 817             velec            = _mm_mul_ps(qq12,rinv12);
 818             felec            = _mm_mul_ps(velec,rinvsq12);
 819
 820             /* Update potential sum for this i atom from the interaction with this j atom. */
 821             velec            = _mm_andnot_ps(dummy_mask,velec);
 822             velecsum         = _mm_add_ps(velecsum,velec);
 823
 824             fscal            = felec;
 825
 826             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 827
 828             /* Calculate temporary vectorial force */
 829             tx               = _mm_mul_ps(fscal,dx12);
 830             ty               = _mm_mul_ps(fscal,dy12);
 831             tz               = _mm_mul_ps(fscal,dz12);
 832
 833             /* Update vectorial force */
 834             fix1             = _mm_add_ps(fix1,tx);
 835             fiy1             = _mm_add_ps(fiy1,ty);
 836             fiz1             = _mm_add_ps(fiz1,tz);
 837
 838             fjx2             = _mm_add_ps(fjx2,tx);
 839             fjy2             = _mm_add_ps(fjy2,ty);
 840             fjz2             = _mm_add_ps(fjz2,tz);
 841
 842             /**************************
 843              * CALCULATE INTERACTIONS *
 844              **************************/
 845
 846             /* COULOMB ELECTROSTATICS */
 847             velec            = _mm_mul_ps(qq20,rinv20);
 848             felec            = _mm_mul_ps(velec,rinvsq20);
 849
 850             /* Update potential sum for this i atom from the interaction with this j atom. */
 851             velec            = _mm_andnot_ps(dummy_mask,velec);
 852             velecsum         = _mm_add_ps(velecsum,velec);
 853
 854             fscal            = felec;
 855
 856             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 857
 858             /* Calculate temporary vectorial force */
 859             tx               = _mm_mul_ps(fscal,dx20);
 860             ty               = _mm_mul_ps(fscal,dy20);
 861             tz               = _mm_mul_ps(fscal,dz20);
 862
 863             /* Update vectorial force */
 864             fix2             = _mm_add_ps(fix2,tx);
 865             fiy2             = _mm_add_ps(fiy2,ty);
 866             fiz2             = _mm_add_ps(fiz2,tz);
 867
 868             fjx0             = _mm_add_ps(fjx0,tx);
 869             fjy0             = _mm_add_ps(fjy0,ty);
 870             fjz0             = _mm_add_ps(fjz0,tz);
 871
 872             /**************************
 873              * CALCULATE INTERACTIONS *
 874              **************************/
 875
 876             /* COULOMB ELECTROSTATICS */
 877             velec            = _mm_mul_ps(qq21,rinv21);
 878             felec            = _mm_mul_ps(velec,rinvsq21);
 879
 880             /* Update potential sum for this i atom from the interaction with this j atom. */
 881             velec            = _mm_andnot_ps(dummy_mask,velec);
 882             velecsum         = _mm_add_ps(velecsum,velec);
 883
 884             fscal            = felec;
 885
 886             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 887
 888             /* Calculate temporary vectorial force */
 889             tx               = _mm_mul_ps(fscal,dx21);
 890             ty               = _mm_mul_ps(fscal,dy21);
 891             tz               = _mm_mul_ps(fscal,dz21);
 892
 893             /* Update vectorial force */
 894             fix2             = _mm_add_ps(fix2,tx);
 895             fiy2             = _mm_add_ps(fiy2,ty);
 896             fiz2             = _mm_add_ps(fiz2,tz);
 897
 898             fjx1             = _mm_add_ps(fjx1,tx);
 899             fjy1             = _mm_add_ps(fjy1,ty);
 900             fjz1             = _mm_add_ps(fjz1,tz);
 901
 902             /**************************
 903              * CALCULATE INTERACTIONS *
 904              **************************/
 905
 906             /* COULOMB ELECTROSTATICS */
 907             velec            = _mm_mul_ps(qq22,rinv22);
 908             felec            = _mm_mul_ps(velec,rinvsq22);
 909
 910             /* Update potential sum for this i atom from the interaction with this j atom. */
 911             velec            = _mm_andnot_ps(dummy_mask,velec);
 912             velecsum         = _mm_add_ps(velecsum,velec);
 913
 914             fscal            = felec;
 915
 916             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 917
 918             /* Calculate temporary vectorial force */
 919             tx               = _mm_mul_ps(fscal,dx22);
 920             ty               = _mm_mul_ps(fscal,dy22);
 921             tz               = _mm_mul_ps(fscal,dz22);
 922
 923             /* Update vectorial force */
 924             fix2             = _mm_add_ps(fix2,tx);
 925             fiy2             = _mm_add_ps(fiy2,ty);
 926             fiz2             = _mm_add_ps(fiz2,tz);
 927
 928             fjx2             = _mm_add_ps(fjx2,tx);
 929             fjy2             = _mm_add_ps(fjy2,ty);
 930             fjz2             = _mm_add_ps(fjz2,tz);
 931
 932             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 933             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 934             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 935             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 936
 937             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 938                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 939
 940             /* Inner loop uses 264 flops */
 941         }
 942
 943         /* End of innermost loop */
 944
 945         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 946                                               f+i_coord_offset,fshift+i_shift_offset);
 947
 948         ggid                        = gid[iidx];
 949         /* Update potential energies */
 950         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 951         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 952
 953         /* Increment number of inner iterations */
 954         inneriter                  += j_index_end - j_index_start;
 955
 956         /* Outer loop uses 20 flops */
 957     }
 958
 959     /* Increment number of outer iterations */
 960     outeriter        += nri;
 961
 962     /* Update outer/inner flops */
 963
 964     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*264);
 965 }
 966 /*
 967  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW3W3_F_sse4_1_single
 968  * Electrostatics interaction: Coulomb
 969  * VdW interaction:            LennardJones
 970  * Geometry:                   Water3-Water3
 971  * Calculate force/pot:        Force
 972  */
 973 void
 974 nb_kernel_ElecCoul_VdwLJ_GeomW3W3_F_sse4_1_single
 975                     (t_nblist                    * gmx_restrict       nlist,
 976                      rvec                        * gmx_restrict          xx,
 977                      rvec                        * gmx_restrict          ff,
 978                      t_forcerec                  * gmx_restrict          fr,
 979                      t_mdatoms                   * gmx_restrict     mdatoms,
 980                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 981                      t_nrnb                      * gmx_restrict        nrnb)
 982 {
 983     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 984      * just 0 for non-waters.
 985      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 986      * jnr indices corresponding to data put in the four positions in the SIMD register.
 987      */
 988     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 989     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 990     int              jnrA,jnrB,jnrC,jnrD;
 991     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 992     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 993     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 994     real             rcutoff_scalar;
 995     real             *shiftvec,*fshift,*x,*f;
 996     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 997     real             scratch[4*DIM];
 998     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 999     int              vdwioffset0;
1000     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1001     int              vdwioffset1;
1002     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1003     int              vdwioffset2;
1004     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1005     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1006     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1007     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1008     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1009     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1010     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1011     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1012     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1013     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1014     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1015     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1016     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1017     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1018     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1019     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1020     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1021     real             *charge;
1022     int              nvdwtype;
1023     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1024     int              *vdwtype;
1025     real             *vdwparam;
1026     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1027     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1028     __m128           dummy_mask,cutoff_mask;
1029     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1030     __m128           one     = _mm_set1_ps(1.0);
1031     __m128           two     = _mm_set1_ps(2.0);
1032     x                = xx[0];
1033     f                = ff[0];
1034
1035     nri              = nlist->nri;
1036     iinr             = nlist->iinr;
1037     jindex           = nlist->jindex;
1038     jjnr             = nlist->jjnr;
1039     shiftidx         = nlist->shift;
1040     gid              = nlist->gid;
1041     shiftvec         = fr->shift_vec[0];
1042     fshift           = fr->fshift[0];
1043     facel            = _mm_set1_ps(fr->epsfac);
1044     charge           = mdatoms->chargeA;
1045     nvdwtype         = fr->ntype;
1046     vdwparam         = fr->nbfp;
1047     vdwtype          = mdatoms->typeA;
1048
1049     /* Setup water-specific parameters */
1050     inr              = nlist->iinr[0];
1051     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1052     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1053     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1054     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1055
1056     jq0              = _mm_set1_ps(charge[inr+0]);
1057     jq1              = _mm_set1_ps(charge[inr+1]);
1058     jq2              = _mm_set1_ps(charge[inr+2]);
1059     vdwjidx0A        = 2*vdwtype[inr+0];
1060     qq00             = _mm_mul_ps(iq0,jq0);
1061     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1062     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1063     qq01             = _mm_mul_ps(iq0,jq1);
1064     qq02             = _mm_mul_ps(iq0,jq2);
1065     qq10             = _mm_mul_ps(iq1,jq0);
1066     qq11             = _mm_mul_ps(iq1,jq1);
1067     qq12             = _mm_mul_ps(iq1,jq2);
1068     qq20             = _mm_mul_ps(iq2,jq0);
1069     qq21             = _mm_mul_ps(iq2,jq1);
1070     qq22             = _mm_mul_ps(iq2,jq2);
1071
1072     /* Avoid stupid compiler warnings */
1073     jnrA = jnrB = jnrC = jnrD = 0;
1074     j_coord_offsetA = 0;
1075     j_coord_offsetB = 0;
1076     j_coord_offsetC = 0;
1077     j_coord_offsetD = 0;
1078
1079     outeriter        = 0;
1080     inneriter        = 0;
1081
1082     for(iidx=0;iidx<4*DIM;iidx++)
1083     {
1084         scratch[iidx] = 0.0;
1085     }
1086
1087     /* Start outer loop over neighborlists */
1088     for(iidx=0; iidx<nri; iidx++)
1089     {
1090         /* Load shift vector for this list */
1091         i_shift_offset   = DIM*shiftidx[iidx];
1092
1093         /* Load limits for loop over neighbors */
1094         j_index_start    = jindex[iidx];
1095         j_index_end      = jindex[iidx+1];
1096
1097         /* Get outer coordinate index */
1098         inr              = iinr[iidx];
1099         i_coord_offset   = DIM*inr;
1100
1101         /* Load i particle coords and add shift vector */
1102         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1103                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1104
1105         fix0             = _mm_setzero_ps();
1106         fiy0             = _mm_setzero_ps();
1107         fiz0             = _mm_setzero_ps();
1108         fix1             = _mm_setzero_ps();
1109         fiy1             = _mm_setzero_ps();
1110         fiz1             = _mm_setzero_ps();
1111         fix2             = _mm_setzero_ps();
1112         fiy2             = _mm_setzero_ps();
1113         fiz2             = _mm_setzero_ps();
1114
1115         /* Start inner kernel loop */
1116         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1117         {
1118
1119             /* Get j neighbor index, and coordinate index */
1120             jnrA             = jjnr[jidx];
1121             jnrB             = jjnr[jidx+1];
1122             jnrC             = jjnr[jidx+2];
1123             jnrD             = jjnr[jidx+3];
1124             j_coord_offsetA  = DIM*jnrA;
1125             j_coord_offsetB  = DIM*jnrB;
1126             j_coord_offsetC  = DIM*jnrC;
1127             j_coord_offsetD  = DIM*jnrD;
1128
1129             /* load j atom coordinates */
1130             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1131                                               x+j_coord_offsetC,x+j_coord_offsetD,
1132                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1133
1134             /* Calculate displacement vector */
1135             dx00             = _mm_sub_ps(ix0,jx0);
1136             dy00             = _mm_sub_ps(iy0,jy0);
1137             dz00             = _mm_sub_ps(iz0,jz0);
1138             dx01             = _mm_sub_ps(ix0,jx1);
1139             dy01             = _mm_sub_ps(iy0,jy1);
1140             dz01             = _mm_sub_ps(iz0,jz1);
1141             dx02             = _mm_sub_ps(ix0,jx2);
1142             dy02             = _mm_sub_ps(iy0,jy2);
1143             dz02             = _mm_sub_ps(iz0,jz2);
1144             dx10             = _mm_sub_ps(ix1,jx0);
1145             dy10             = _mm_sub_ps(iy1,jy0);
1146             dz10             = _mm_sub_ps(iz1,jz0);
1147             dx11             = _mm_sub_ps(ix1,jx1);
1148             dy11             = _mm_sub_ps(iy1,jy1);
1149             dz11             = _mm_sub_ps(iz1,jz1);
1150             dx12             = _mm_sub_ps(ix1,jx2);
1151             dy12             = _mm_sub_ps(iy1,jy2);
1152             dz12             = _mm_sub_ps(iz1,jz2);
1153             dx20             = _mm_sub_ps(ix2,jx0);
1154             dy20             = _mm_sub_ps(iy2,jy0);
1155             dz20             = _mm_sub_ps(iz2,jz0);
1156             dx21             = _mm_sub_ps(ix2,jx1);
1157             dy21             = _mm_sub_ps(iy2,jy1);
1158             dz21             = _mm_sub_ps(iz2,jz1);
1159             dx22             = _mm_sub_ps(ix2,jx2);
1160             dy22             = _mm_sub_ps(iy2,jy2);
1161             dz22             = _mm_sub_ps(iz2,jz2);
1162
1163             /* Calculate squared distance and things based on it */
1164             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1165             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1166             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1167             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1168             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1169             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1170             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1171             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1172             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1173
1174             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1175             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1176             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1177             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1178             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1179             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1180             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1181             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1182             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1183
1184             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1185             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1186             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1187             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1188             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1189             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1190             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1191             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1192             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1193
1194             fjx0             = _mm_setzero_ps();
1195             fjy0             = _mm_setzero_ps();
1196             fjz0             = _mm_setzero_ps();
1197             fjx1             = _mm_setzero_ps();
1198             fjy1             = _mm_setzero_ps();
1199             fjz1             = _mm_setzero_ps();
1200             fjx2             = _mm_setzero_ps();
1201             fjy2             = _mm_setzero_ps();
1202             fjz2             = _mm_setzero_ps();
1203
1204             /**************************
1205              * CALCULATE INTERACTIONS *
1206              **************************/
1207
1208             /* COULOMB ELECTROSTATICS */
1209             velec            = _mm_mul_ps(qq00,rinv00);
1210             felec            = _mm_mul_ps(velec,rinvsq00);
1211
1212             /* LENNARD-JONES DISPERSION/REPULSION */
1213
1214             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1215             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1216
1217             fscal            = _mm_add_ps(felec,fvdw);
1218
1219             /* Calculate temporary vectorial force */
1220             tx               = _mm_mul_ps(fscal,dx00);
1221             ty               = _mm_mul_ps(fscal,dy00);
1222             tz               = _mm_mul_ps(fscal,dz00);
1223
1224             /* Update vectorial force */
1225             fix0             = _mm_add_ps(fix0,tx);
1226             fiy0             = _mm_add_ps(fiy0,ty);
1227             fiz0             = _mm_add_ps(fiz0,tz);
1228
1229             fjx0             = _mm_add_ps(fjx0,tx);
1230             fjy0             = _mm_add_ps(fjy0,ty);
1231             fjz0             = _mm_add_ps(fjz0,tz);
1232
1233             /**************************
1234              * CALCULATE INTERACTIONS *
1235              **************************/
1236
1237             /* COULOMB ELECTROSTATICS */
1238             velec            = _mm_mul_ps(qq01,rinv01);
1239             felec            = _mm_mul_ps(velec,rinvsq01);
1240
1241             fscal            = felec;
1242
1243             /* Calculate temporary vectorial force */
1244             tx               = _mm_mul_ps(fscal,dx01);
1245             ty               = _mm_mul_ps(fscal,dy01);
1246             tz               = _mm_mul_ps(fscal,dz01);
1247
1248             /* Update vectorial force */
1249             fix0             = _mm_add_ps(fix0,tx);
1250             fiy0             = _mm_add_ps(fiy0,ty);
1251             fiz0             = _mm_add_ps(fiz0,tz);
1252
1253             fjx1             = _mm_add_ps(fjx1,tx);
1254             fjy1             = _mm_add_ps(fjy1,ty);
1255             fjz1             = _mm_add_ps(fjz1,tz);
1256
1257             /**************************
1258              * CALCULATE INTERACTIONS *
1259              **************************/
1260
1261             /* COULOMB ELECTROSTATICS */
1262             velec            = _mm_mul_ps(qq02,rinv02);
1263             felec            = _mm_mul_ps(velec,rinvsq02);
1264
1265             fscal            = felec;
1266
1267             /* Calculate temporary vectorial force */
1268             tx               = _mm_mul_ps(fscal,dx02);
1269             ty               = _mm_mul_ps(fscal,dy02);
1270             tz               = _mm_mul_ps(fscal,dz02);
1271
1272             /* Update vectorial force */
1273             fix0             = _mm_add_ps(fix0,tx);
1274             fiy0             = _mm_add_ps(fiy0,ty);
1275             fiz0             = _mm_add_ps(fiz0,tz);
1276
1277             fjx2             = _mm_add_ps(fjx2,tx);
1278             fjy2             = _mm_add_ps(fjy2,ty);
1279             fjz2             = _mm_add_ps(fjz2,tz);
1280
1281             /**************************
1282              * CALCULATE INTERACTIONS *
1283              **************************/
1284
1285             /* COULOMB ELECTROSTATICS */
1286             velec            = _mm_mul_ps(qq10,rinv10);
1287             felec            = _mm_mul_ps(velec,rinvsq10);
1288
1289             fscal            = felec;
1290
1291             /* Calculate temporary vectorial force */
1292             tx               = _mm_mul_ps(fscal,dx10);
1293             ty               = _mm_mul_ps(fscal,dy10);
1294             tz               = _mm_mul_ps(fscal,dz10);
1295
1296             /* Update vectorial force */
1297             fix1             = _mm_add_ps(fix1,tx);
1298             fiy1             = _mm_add_ps(fiy1,ty);
1299             fiz1             = _mm_add_ps(fiz1,tz);
1300
1301             fjx0             = _mm_add_ps(fjx0,tx);
1302             fjy0             = _mm_add_ps(fjy0,ty);
1303             fjz0             = _mm_add_ps(fjz0,tz);
1304
1305             /**************************
1306              * CALCULATE INTERACTIONS *
1307              **************************/
1308
1309             /* COULOMB ELECTROSTATICS */
1310             velec            = _mm_mul_ps(qq11,rinv11);
1311             felec            = _mm_mul_ps(velec,rinvsq11);
1312
1313             fscal            = felec;
1314
1315             /* Calculate temporary vectorial force */
1316             tx               = _mm_mul_ps(fscal,dx11);
1317             ty               = _mm_mul_ps(fscal,dy11);
1318             tz               = _mm_mul_ps(fscal,dz11);
1319
1320             /* Update vectorial force */
1321             fix1             = _mm_add_ps(fix1,tx);
1322             fiy1             = _mm_add_ps(fiy1,ty);
1323             fiz1             = _mm_add_ps(fiz1,tz);
1324
1325             fjx1             = _mm_add_ps(fjx1,tx);
1326             fjy1             = _mm_add_ps(fjy1,ty);
1327             fjz1             = _mm_add_ps(fjz1,tz);
1328
1329             /**************************
1330              * CALCULATE INTERACTIONS *
1331              **************************/
1332
1333             /* COULOMB ELECTROSTATICS */
1334             velec            = _mm_mul_ps(qq12,rinv12);
1335             felec            = _mm_mul_ps(velec,rinvsq12);
1336
1337             fscal            = felec;
1338
1339             /* Calculate temporary vectorial force */
1340             tx               = _mm_mul_ps(fscal,dx12);
1341             ty               = _mm_mul_ps(fscal,dy12);
1342             tz               = _mm_mul_ps(fscal,dz12);
1343
1344             /* Update vectorial force */
1345             fix1             = _mm_add_ps(fix1,tx);
1346             fiy1             = _mm_add_ps(fiy1,ty);
1347             fiz1             = _mm_add_ps(fiz1,tz);
1348
1349             fjx2             = _mm_add_ps(fjx2,tx);
1350             fjy2             = _mm_add_ps(fjy2,ty);
1351             fjz2             = _mm_add_ps(fjz2,tz);
1352
1353             /**************************
1354              * CALCULATE INTERACTIONS *
1355              **************************/
1356
1357             /* COULOMB ELECTROSTATICS */
1358             velec            = _mm_mul_ps(qq20,rinv20);
1359             felec            = _mm_mul_ps(velec,rinvsq20);
1360
1361             fscal            = felec;
1362
1363             /* Calculate temporary vectorial force */
1364             tx               = _mm_mul_ps(fscal,dx20);
1365             ty               = _mm_mul_ps(fscal,dy20);
1366             tz               = _mm_mul_ps(fscal,dz20);
1367
1368             /* Update vectorial force */
1369             fix2             = _mm_add_ps(fix2,tx);
1370             fiy2             = _mm_add_ps(fiy2,ty);
1371             fiz2             = _mm_add_ps(fiz2,tz);
1372
1373             fjx0             = _mm_add_ps(fjx0,tx);
1374             fjy0             = _mm_add_ps(fjy0,ty);
1375             fjz0             = _mm_add_ps(fjz0,tz);
1376
1377             /**************************
1378              * CALCULATE INTERACTIONS *
1379              **************************/
1380
1381             /* COULOMB ELECTROSTATICS */
1382             velec            = _mm_mul_ps(qq21,rinv21);
1383             felec            = _mm_mul_ps(velec,rinvsq21);
1384
1385             fscal            = felec;
1386
1387             /* Calculate temporary vectorial force */
1388             tx               = _mm_mul_ps(fscal,dx21);
1389             ty               = _mm_mul_ps(fscal,dy21);
1390             tz               = _mm_mul_ps(fscal,dz21);
1391
1392             /* Update vectorial force */
1393             fix2             = _mm_add_ps(fix2,tx);
1394             fiy2             = _mm_add_ps(fiy2,ty);
1395             fiz2             = _mm_add_ps(fiz2,tz);
1396
1397             fjx1             = _mm_add_ps(fjx1,tx);
1398             fjy1             = _mm_add_ps(fjy1,ty);
1399             fjz1             = _mm_add_ps(fjz1,tz);
1400
1401             /**************************
1402              * CALCULATE INTERACTIONS *
1403              **************************/
1404
1405             /* COULOMB ELECTROSTATICS */
1406             velec            = _mm_mul_ps(qq22,rinv22);
1407             felec            = _mm_mul_ps(velec,rinvsq22);
1408
1409             fscal            = felec;
1410
1411             /* Calculate temporary vectorial force */
1412             tx               = _mm_mul_ps(fscal,dx22);
1413             ty               = _mm_mul_ps(fscal,dy22);
1414             tz               = _mm_mul_ps(fscal,dz22);
1415
1416             /* Update vectorial force */
1417             fix2             = _mm_add_ps(fix2,tx);
1418             fiy2             = _mm_add_ps(fiy2,ty);
1419             fiz2             = _mm_add_ps(fiz2,tz);
1420
1421             fjx2             = _mm_add_ps(fjx2,tx);
1422             fjy2             = _mm_add_ps(fjy2,ty);
1423             fjz2             = _mm_add_ps(fjz2,tz);
1424
1425             fjptrA             = f+j_coord_offsetA;
1426             fjptrB             = f+j_coord_offsetB;
1427             fjptrC             = f+j_coord_offsetC;
1428             fjptrD             = f+j_coord_offsetD;
1429
1430             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1431                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1432
1433             /* Inner loop uses 250 flops */
1434         }
1435
1436         if(jidx<j_index_end)
1437         {
1438
1439             /* Get j neighbor index, and coordinate index */
1440             jnrlistA         = jjnr[jidx];
1441             jnrlistB         = jjnr[jidx+1];
1442             jnrlistC         = jjnr[jidx+2];
1443             jnrlistD         = jjnr[jidx+3];
1444             /* Sign of each element will be negative for non-real atoms.
1445              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1446              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1447              */
1448             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1449             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1450             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1451             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1452             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1453             j_coord_offsetA  = DIM*jnrA;
1454             j_coord_offsetB  = DIM*jnrB;
1455             j_coord_offsetC  = DIM*jnrC;
1456             j_coord_offsetD  = DIM*jnrD;
1457
1458             /* load j atom coordinates */
1459             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1460                                               x+j_coord_offsetC,x+j_coord_offsetD,
1461                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1462
1463             /* Calculate displacement vector */
1464             dx00             = _mm_sub_ps(ix0,jx0);
1465             dy00             = _mm_sub_ps(iy0,jy0);
1466             dz00             = _mm_sub_ps(iz0,jz0);
1467             dx01             = _mm_sub_ps(ix0,jx1);
1468             dy01             = _mm_sub_ps(iy0,jy1);
1469             dz01             = _mm_sub_ps(iz0,jz1);
1470             dx02             = _mm_sub_ps(ix0,jx2);
1471             dy02             = _mm_sub_ps(iy0,jy2);
1472             dz02             = _mm_sub_ps(iz0,jz2);
1473             dx10             = _mm_sub_ps(ix1,jx0);
1474             dy10             = _mm_sub_ps(iy1,jy0);
1475             dz10             = _mm_sub_ps(iz1,jz0);
1476             dx11             = _mm_sub_ps(ix1,jx1);
1477             dy11             = _mm_sub_ps(iy1,jy1);
1478             dz11             = _mm_sub_ps(iz1,jz1);
1479             dx12             = _mm_sub_ps(ix1,jx2);
1480             dy12             = _mm_sub_ps(iy1,jy2);
1481             dz12             = _mm_sub_ps(iz1,jz2);
1482             dx20             = _mm_sub_ps(ix2,jx0);
1483             dy20             = _mm_sub_ps(iy2,jy0);
1484             dz20             = _mm_sub_ps(iz2,jz0);
1485             dx21             = _mm_sub_ps(ix2,jx1);
1486             dy21             = _mm_sub_ps(iy2,jy1);
1487             dz21             = _mm_sub_ps(iz2,jz1);
1488             dx22             = _mm_sub_ps(ix2,jx2);
1489             dy22             = _mm_sub_ps(iy2,jy2);
1490             dz22             = _mm_sub_ps(iz2,jz2);
1491
1492             /* Calculate squared distance and things based on it */
1493             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1494             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1495             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1496             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1497             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1498             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1499             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1500             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1501             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1502
1503             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1504             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1505             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1506             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1507             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1508             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1509             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1510             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1511             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1512
1513             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1514             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1515             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1516             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1517             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1518             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1519             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1520             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1521             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1522
1523             fjx0             = _mm_setzero_ps();
1524             fjy0             = _mm_setzero_ps();
1525             fjz0             = _mm_setzero_ps();
1526             fjx1             = _mm_setzero_ps();
1527             fjy1             = _mm_setzero_ps();
1528             fjz1             = _mm_setzero_ps();
1529             fjx2             = _mm_setzero_ps();
1530             fjy2             = _mm_setzero_ps();
1531             fjz2             = _mm_setzero_ps();
1532
1533             /**************************
1534              * CALCULATE INTERACTIONS *
1535              **************************/
1536
1537             /* COULOMB ELECTROSTATICS */
1538             velec            = _mm_mul_ps(qq00,rinv00);
1539             felec            = _mm_mul_ps(velec,rinvsq00);
1540
1541             /* LENNARD-JONES DISPERSION/REPULSION */
1542
1543             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1544             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1545
1546             fscal            = _mm_add_ps(felec,fvdw);
1547
1548             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1549
1550             /* Calculate temporary vectorial force */
1551             tx               = _mm_mul_ps(fscal,dx00);
1552             ty               = _mm_mul_ps(fscal,dy00);
1553             tz               = _mm_mul_ps(fscal,dz00);
1554
1555             /* Update vectorial force */
1556             fix0             = _mm_add_ps(fix0,tx);
1557             fiy0             = _mm_add_ps(fiy0,ty);
1558             fiz0             = _mm_add_ps(fiz0,tz);
1559
1560             fjx0             = _mm_add_ps(fjx0,tx);
1561             fjy0             = _mm_add_ps(fjy0,ty);
1562             fjz0             = _mm_add_ps(fjz0,tz);
1563
1564             /**************************
1565              * CALCULATE INTERACTIONS *
1566              **************************/
1567
1568             /* COULOMB ELECTROSTATICS */
1569             velec            = _mm_mul_ps(qq01,rinv01);
1570             felec            = _mm_mul_ps(velec,rinvsq01);
1571
1572             fscal            = felec;
1573
1574             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1575
1576             /* Calculate temporary vectorial force */
1577             tx               = _mm_mul_ps(fscal,dx01);
1578             ty               = _mm_mul_ps(fscal,dy01);
1579             tz               = _mm_mul_ps(fscal,dz01);
1580
1581             /* Update vectorial force */
1582             fix0             = _mm_add_ps(fix0,tx);
1583             fiy0             = _mm_add_ps(fiy0,ty);
1584             fiz0             = _mm_add_ps(fiz0,tz);
1585
1586             fjx1             = _mm_add_ps(fjx1,tx);
1587             fjy1             = _mm_add_ps(fjy1,ty);
1588             fjz1             = _mm_add_ps(fjz1,tz);
1589
1590             /**************************
1591              * CALCULATE INTERACTIONS *
1592              **************************/
1593
1594             /* COULOMB ELECTROSTATICS */
1595             velec            = _mm_mul_ps(qq02,rinv02);
1596             felec            = _mm_mul_ps(velec,rinvsq02);
1597
1598             fscal            = felec;
1599
1600             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1601
1602             /* Calculate temporary vectorial force */
1603             tx               = _mm_mul_ps(fscal,dx02);
1604             ty               = _mm_mul_ps(fscal,dy02);
1605             tz               = _mm_mul_ps(fscal,dz02);
1606
1607             /* Update vectorial force */
1608             fix0             = _mm_add_ps(fix0,tx);
1609             fiy0             = _mm_add_ps(fiy0,ty);
1610             fiz0             = _mm_add_ps(fiz0,tz);
1611
1612             fjx2             = _mm_add_ps(fjx2,tx);
1613             fjy2             = _mm_add_ps(fjy2,ty);
1614             fjz2             = _mm_add_ps(fjz2,tz);
1615
1616             /**************************
1617              * CALCULATE INTERACTIONS *
1618              **************************/
1619
1620             /* COULOMB ELECTROSTATICS */
1621             velec            = _mm_mul_ps(qq10,rinv10);
1622             felec            = _mm_mul_ps(velec,rinvsq10);
1623
1624             fscal            = felec;
1625
1626             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1627
1628             /* Calculate temporary vectorial force */
1629             tx               = _mm_mul_ps(fscal,dx10);
1630             ty               = _mm_mul_ps(fscal,dy10);
1631             tz               = _mm_mul_ps(fscal,dz10);
1632
1633             /* Update vectorial force */
1634             fix1             = _mm_add_ps(fix1,tx);
1635             fiy1             = _mm_add_ps(fiy1,ty);
1636             fiz1             = _mm_add_ps(fiz1,tz);
1637
1638             fjx0             = _mm_add_ps(fjx0,tx);
1639             fjy0             = _mm_add_ps(fjy0,ty);
1640             fjz0             = _mm_add_ps(fjz0,tz);
1641
1642             /**************************
1643              * CALCULATE INTERACTIONS *
1644              **************************/
1645
1646             /* COULOMB ELECTROSTATICS */
1647             velec            = _mm_mul_ps(qq11,rinv11);
1648             felec            = _mm_mul_ps(velec,rinvsq11);
1649
1650             fscal            = felec;
1651
1652             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1653
1654             /* Calculate temporary vectorial force */
1655             tx               = _mm_mul_ps(fscal,dx11);
1656             ty               = _mm_mul_ps(fscal,dy11);
1657             tz               = _mm_mul_ps(fscal,dz11);
1658
1659             /* Update vectorial force */
1660             fix1             = _mm_add_ps(fix1,tx);
1661             fiy1             = _mm_add_ps(fiy1,ty);
1662             fiz1             = _mm_add_ps(fiz1,tz);
1663
1664             fjx1             = _mm_add_ps(fjx1,tx);
1665             fjy1             = _mm_add_ps(fjy1,ty);
1666             fjz1             = _mm_add_ps(fjz1,tz);
1667
1668             /**************************
1669              * CALCULATE INTERACTIONS *
1670              **************************/
1671
1672             /* COULOMB ELECTROSTATICS */
1673             velec            = _mm_mul_ps(qq12,rinv12);
1674             felec            = _mm_mul_ps(velec,rinvsq12);
1675
1676             fscal            = felec;
1677
1678             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1679
1680             /* Calculate temporary vectorial force */
1681             tx               = _mm_mul_ps(fscal,dx12);
1682             ty               = _mm_mul_ps(fscal,dy12);
1683             tz               = _mm_mul_ps(fscal,dz12);
1684
1685             /* Update vectorial force */
1686             fix1             = _mm_add_ps(fix1,tx);
1687             fiy1             = _mm_add_ps(fiy1,ty);
1688             fiz1             = _mm_add_ps(fiz1,tz);
1689
1690             fjx2             = _mm_add_ps(fjx2,tx);
1691             fjy2             = _mm_add_ps(fjy2,ty);
1692             fjz2             = _mm_add_ps(fjz2,tz);
1693
1694             /**************************
1695              * CALCULATE INTERACTIONS *
1696              **************************/
1697
1698             /* COULOMB ELECTROSTATICS */
1699             velec            = _mm_mul_ps(qq20,rinv20);
1700             felec            = _mm_mul_ps(velec,rinvsq20);
1701
1702             fscal            = felec;
1703
1704             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1705
1706             /* Calculate temporary vectorial force */
1707             tx               = _mm_mul_ps(fscal,dx20);
1708             ty               = _mm_mul_ps(fscal,dy20);
1709             tz               = _mm_mul_ps(fscal,dz20);
1710
1711             /* Update vectorial force */
1712             fix2             = _mm_add_ps(fix2,tx);
1713             fiy2             = _mm_add_ps(fiy2,ty);
1714             fiz2             = _mm_add_ps(fiz2,tz);
1715
1716             fjx0             = _mm_add_ps(fjx0,tx);
1717             fjy0             = _mm_add_ps(fjy0,ty);
1718             fjz0             = _mm_add_ps(fjz0,tz);
1719
1720             /**************************
1721              * CALCULATE INTERACTIONS *
1722              **************************/
1723
1724             /* COULOMB ELECTROSTATICS */
1725             velec            = _mm_mul_ps(qq21,rinv21);
1726             felec            = _mm_mul_ps(velec,rinvsq21);
1727
1728             fscal            = felec;
1729
1730             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1731
1732             /* Calculate temporary vectorial force */
1733             tx               = _mm_mul_ps(fscal,dx21);
1734             ty               = _mm_mul_ps(fscal,dy21);
1735             tz               = _mm_mul_ps(fscal,dz21);
1736
1737             /* Update vectorial force */
1738             fix2             = _mm_add_ps(fix2,tx);
1739             fiy2             = _mm_add_ps(fiy2,ty);
1740             fiz2             = _mm_add_ps(fiz2,tz);
1741
1742             fjx1             = _mm_add_ps(fjx1,tx);
1743             fjy1             = _mm_add_ps(fjy1,ty);
1744             fjz1             = _mm_add_ps(fjz1,tz);
1745
1746             /**************************
1747              * CALCULATE INTERACTIONS *
1748              **************************/
1749
1750             /* COULOMB ELECTROSTATICS */
1751             velec            = _mm_mul_ps(qq22,rinv22);
1752             felec            = _mm_mul_ps(velec,rinvsq22);
1753
1754             fscal            = felec;
1755
1756             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1757
1758             /* Calculate temporary vectorial force */
1759             tx               = _mm_mul_ps(fscal,dx22);
1760             ty               = _mm_mul_ps(fscal,dy22);
1761             tz               = _mm_mul_ps(fscal,dz22);
1762
1763             /* Update vectorial force */
1764             fix2             = _mm_add_ps(fix2,tx);
1765             fiy2             = _mm_add_ps(fiy2,ty);
1766             fiz2             = _mm_add_ps(fiz2,tz);
1767
1768             fjx2             = _mm_add_ps(fjx2,tx);
1769             fjy2             = _mm_add_ps(fjy2,ty);
1770             fjz2             = _mm_add_ps(fjz2,tz);
1771
1772             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1773             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1774             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1775             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1776
1777             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1778                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1779
1780             /* Inner loop uses 250 flops */
1781         }
1782
1783         /* End of innermost loop */
1784
1785         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1786                                               f+i_coord_offset,fshift+i_shift_offset);
1787
1788         /* Increment number of inner iterations */
1789         inneriter                  += j_index_end - j_index_start;
1790
1791         /* Outer loop uses 18 flops */
1792     }
1793
1794     /* Increment number of outer iterations */
1795     outeriter        += nri;
1796
1797     /* Update outer/inner flops */
1798
1799     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*250);
1800 }