src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_ElecCoul_VdwLJ_GeomW3W3_sse4_1_single.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, 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_GeomW3W3_VF_sse4_1_single
  51  * Electrostatics interaction: Coulomb
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Water3-Water3
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecCoul_VdwLJ_GeomW3W3_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              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  89     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  91     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  92     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  93     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  94     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  95     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  96     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  97     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  98     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  99     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 100     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 101     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 102     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 103     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 104     real             *charge;
 105     int              nvdwtype;
 106     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 107     int              *vdwtype;
 108     real             *vdwparam;
 109     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 110     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 111     __m128           dummy_mask,cutoff_mask;
 112     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 113     __m128           one     = _mm_set1_ps(1.0);
 114     __m128           two     = _mm_set1_ps(2.0);
 115     x                = xx[0];
 116     f                = ff[0];
 117
 118     nri              = nlist->nri;
 119     iinr             = nlist->iinr;
 120     jindex           = nlist->jindex;
 121     jjnr             = nlist->jjnr;
 122     shiftidx         = nlist->shift;
 123     gid              = nlist->gid;
 124     shiftvec         = fr->shift_vec[0];
 125     fshift           = fr->fshift[0];
 126     facel            = _mm_set1_ps(fr->ic->epsfac);
 127     charge           = mdatoms->chargeA;
 128     nvdwtype         = fr->ntype;
 129     vdwparam         = fr->nbfp;
 130     vdwtype          = mdatoms->typeA;
 131
 132     /* Setup water-specific parameters */
 133     inr              = nlist->iinr[0];
 134     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 135     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 136     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 137     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 138
 139     jq0              = _mm_set1_ps(charge[inr+0]);
 140     jq1              = _mm_set1_ps(charge[inr+1]);
 141     jq2              = _mm_set1_ps(charge[inr+2]);
 142     vdwjidx0A        = 2*vdwtype[inr+0];
 143     qq00             = _mm_mul_ps(iq0,jq0);
 144     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 145     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 146     qq01             = _mm_mul_ps(iq0,jq1);
 147     qq02             = _mm_mul_ps(iq0,jq2);
 148     qq10             = _mm_mul_ps(iq1,jq0);
 149     qq11             = _mm_mul_ps(iq1,jq1);
 150     qq12             = _mm_mul_ps(iq1,jq2);
 151     qq20             = _mm_mul_ps(iq2,jq0);
 152     qq21             = _mm_mul_ps(iq2,jq1);
 153     qq22             = _mm_mul_ps(iq2,jq2);
 154
 155     /* Avoid stupid compiler warnings */
 156     jnrA = jnrB = jnrC = jnrD = 0;
 157     j_coord_offsetA = 0;
 158     j_coord_offsetB = 0;
 159     j_coord_offsetC = 0;
 160     j_coord_offsetD = 0;
 161
 162     outeriter        = 0;
 163     inneriter        = 0;
 164
 165     for(iidx=0;iidx<4*DIM;iidx++)
 166     {
 167         scratch[iidx] = 0.0;
 168     }
 169
 170     /* Start outer loop over neighborlists */
 171     for(iidx=0; iidx<nri; iidx++)
 172     {
 173         /* Load shift vector for this list */
 174         i_shift_offset   = DIM*shiftidx[iidx];
 175
 176         /* Load limits for loop over neighbors */
 177         j_index_start    = jindex[iidx];
 178         j_index_end      = jindex[iidx+1];
 179
 180         /* Get outer coordinate index */
 181         inr              = iinr[iidx];
 182         i_coord_offset   = DIM*inr;
 183
 184         /* Load i particle coords and add shift vector */
 185         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 186                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 187
 188         fix0             = _mm_setzero_ps();
 189         fiy0             = _mm_setzero_ps();
 190         fiz0             = _mm_setzero_ps();
 191         fix1             = _mm_setzero_ps();
 192         fiy1             = _mm_setzero_ps();
 193         fiz1             = _mm_setzero_ps();
 194         fix2             = _mm_setzero_ps();
 195         fiy2             = _mm_setzero_ps();
 196         fiz2             = _mm_setzero_ps();
 197
 198         /* Reset potential sums */
 199         velecsum         = _mm_setzero_ps();
 200         vvdwsum          = _mm_setzero_ps();
 201
 202         /* Start inner kernel loop */
 203         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 204         {
 205
 206             /* Get j neighbor index, and coordinate index */
 207             jnrA             = jjnr[jidx];
 208             jnrB             = jjnr[jidx+1];
 209             jnrC             = jjnr[jidx+2];
 210             jnrD             = jjnr[jidx+3];
 211             j_coord_offsetA  = DIM*jnrA;
 212             j_coord_offsetB  = DIM*jnrB;
 213             j_coord_offsetC  = DIM*jnrC;
 214             j_coord_offsetD  = DIM*jnrD;
 215
 216             /* load j atom coordinates */
 217             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 218                                               x+j_coord_offsetC,x+j_coord_offsetD,
 219                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 220
 221             /* Calculate displacement vector */
 222             dx00             = _mm_sub_ps(ix0,jx0);
 223             dy00             = _mm_sub_ps(iy0,jy0);
 224             dz00             = _mm_sub_ps(iz0,jz0);
 225             dx01             = _mm_sub_ps(ix0,jx1);
 226             dy01             = _mm_sub_ps(iy0,jy1);
 227             dz01             = _mm_sub_ps(iz0,jz1);
 228             dx02             = _mm_sub_ps(ix0,jx2);
 229             dy02             = _mm_sub_ps(iy0,jy2);
 230             dz02             = _mm_sub_ps(iz0,jz2);
 231             dx10             = _mm_sub_ps(ix1,jx0);
 232             dy10             = _mm_sub_ps(iy1,jy0);
 233             dz10             = _mm_sub_ps(iz1,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             dx20             = _mm_sub_ps(ix2,jx0);
 241             dy20             = _mm_sub_ps(iy2,jy0);
 242             dz20             = _mm_sub_ps(iz2,jz0);
 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
 250             /* Calculate squared distance and things based on it */
 251             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 252             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 253             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 254             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 255             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 256             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 257             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 258             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 259             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 260
 261             rinv00           = sse41_invsqrt_f(rsq00);
 262             rinv01           = sse41_invsqrt_f(rsq01);
 263             rinv02           = sse41_invsqrt_f(rsq02);
 264             rinv10           = sse41_invsqrt_f(rsq10);
 265             rinv11           = sse41_invsqrt_f(rsq11);
 266             rinv12           = sse41_invsqrt_f(rsq12);
 267             rinv20           = sse41_invsqrt_f(rsq20);
 268             rinv21           = sse41_invsqrt_f(rsq21);
 269             rinv22           = sse41_invsqrt_f(rsq22);
 270
 271             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 272             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 273             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 274             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 275             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 276             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 277             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 278             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 279             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 280
 281             fjx0             = _mm_setzero_ps();
 282             fjy0             = _mm_setzero_ps();
 283             fjz0             = _mm_setzero_ps();
 284             fjx1             = _mm_setzero_ps();
 285             fjy1             = _mm_setzero_ps();
 286             fjz1             = _mm_setzero_ps();
 287             fjx2             = _mm_setzero_ps();
 288             fjy2             = _mm_setzero_ps();
 289             fjz2             = _mm_setzero_ps();
 290
 291             /**************************
 292              * CALCULATE INTERACTIONS *
 293              **************************/
 294
 295             /* COULOMB ELECTROSTATICS */
 296             velec            = _mm_mul_ps(qq00,rinv00);
 297             felec            = _mm_mul_ps(velec,rinvsq00);
 298
 299             /* LENNARD-JONES DISPERSION/REPULSION */
 300
 301             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 302             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 303             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 304             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 305             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 306
 307             /* Update potential sum for this i atom from the interaction with this j atom. */
 308             velecsum         = _mm_add_ps(velecsum,velec);
 309             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 310
 311             fscal            = _mm_add_ps(felec,fvdw);
 312
 313             /* Calculate temporary vectorial force */
 314             tx               = _mm_mul_ps(fscal,dx00);
 315             ty               = _mm_mul_ps(fscal,dy00);
 316             tz               = _mm_mul_ps(fscal,dz00);
 317
 318             /* Update vectorial force */
 319             fix0             = _mm_add_ps(fix0,tx);
 320             fiy0             = _mm_add_ps(fiy0,ty);
 321             fiz0             = _mm_add_ps(fiz0,tz);
 322
 323             fjx0             = _mm_add_ps(fjx0,tx);
 324             fjy0             = _mm_add_ps(fjy0,ty);
 325             fjz0             = _mm_add_ps(fjz0,tz);
 326
 327             /**************************
 328              * CALCULATE INTERACTIONS *
 329              **************************/
 330
 331             /* COULOMB ELECTROSTATICS */
 332             velec            = _mm_mul_ps(qq01,rinv01);
 333             felec            = _mm_mul_ps(velec,rinvsq01);
 334
 335             /* Update potential sum for this i atom from the interaction with this j atom. */
 336             velecsum         = _mm_add_ps(velecsum,velec);
 337
 338             fscal            = felec;
 339
 340             /* Calculate temporary vectorial force */
 341             tx               = _mm_mul_ps(fscal,dx01);
 342             ty               = _mm_mul_ps(fscal,dy01);
 343             tz               = _mm_mul_ps(fscal,dz01);
 344
 345             /* Update vectorial force */
 346             fix0             = _mm_add_ps(fix0,tx);
 347             fiy0             = _mm_add_ps(fiy0,ty);
 348             fiz0             = _mm_add_ps(fiz0,tz);
 349
 350             fjx1             = _mm_add_ps(fjx1,tx);
 351             fjy1             = _mm_add_ps(fjy1,ty);
 352             fjz1             = _mm_add_ps(fjz1,tz);
 353
 354             /**************************
 355              * CALCULATE INTERACTIONS *
 356              **************************/
 357
 358             /* COULOMB ELECTROSTATICS */
 359             velec            = _mm_mul_ps(qq02,rinv02);
 360             felec            = _mm_mul_ps(velec,rinvsq02);
 361
 362             /* Update potential sum for this i atom from the interaction with this j atom. */
 363             velecsum         = _mm_add_ps(velecsum,velec);
 364
 365             fscal            = felec;
 366
 367             /* Calculate temporary vectorial force */
 368             tx               = _mm_mul_ps(fscal,dx02);
 369             ty               = _mm_mul_ps(fscal,dy02);
 370             tz               = _mm_mul_ps(fscal,dz02);
 371
 372             /* Update vectorial force */
 373             fix0             = _mm_add_ps(fix0,tx);
 374             fiy0             = _mm_add_ps(fiy0,ty);
 375             fiz0             = _mm_add_ps(fiz0,tz);
 376
 377             fjx2             = _mm_add_ps(fjx2,tx);
 378             fjy2             = _mm_add_ps(fjy2,ty);
 379             fjz2             = _mm_add_ps(fjz2,tz);
 380
 381             /**************************
 382              * CALCULATE INTERACTIONS *
 383              **************************/
 384
 385             /* COULOMB ELECTROSTATICS */
 386             velec            = _mm_mul_ps(qq10,rinv10);
 387             felec            = _mm_mul_ps(velec,rinvsq10);
 388
 389             /* Update potential sum for this i atom from the interaction with this j atom. */
 390             velecsum         = _mm_add_ps(velecsum,velec);
 391
 392             fscal            = felec;
 393
 394             /* Calculate temporary vectorial force */
 395             tx               = _mm_mul_ps(fscal,dx10);
 396             ty               = _mm_mul_ps(fscal,dy10);
 397             tz               = _mm_mul_ps(fscal,dz10);
 398
 399             /* Update vectorial force */
 400             fix1             = _mm_add_ps(fix1,tx);
 401             fiy1             = _mm_add_ps(fiy1,ty);
 402             fiz1             = _mm_add_ps(fiz1,tz);
 403
 404             fjx0             = _mm_add_ps(fjx0,tx);
 405             fjy0             = _mm_add_ps(fjy0,ty);
 406             fjz0             = _mm_add_ps(fjz0,tz);
 407
 408             /**************************
 409              * CALCULATE INTERACTIONS *
 410              **************************/
 411
 412             /* COULOMB ELECTROSTATICS */
 413             velec            = _mm_mul_ps(qq11,rinv11);
 414             felec            = _mm_mul_ps(velec,rinvsq11);
 415
 416             /* Update potential sum for this i atom from the interaction with this j atom. */
 417             velecsum         = _mm_add_ps(velecsum,velec);
 418
 419             fscal            = felec;
 420
 421             /* Calculate temporary vectorial force */
 422             tx               = _mm_mul_ps(fscal,dx11);
 423             ty               = _mm_mul_ps(fscal,dy11);
 424             tz               = _mm_mul_ps(fscal,dz11);
 425
 426             /* Update vectorial force */
 427             fix1             = _mm_add_ps(fix1,tx);
 428             fiy1             = _mm_add_ps(fiy1,ty);
 429             fiz1             = _mm_add_ps(fiz1,tz);
 430
 431             fjx1             = _mm_add_ps(fjx1,tx);
 432             fjy1             = _mm_add_ps(fjy1,ty);
 433             fjz1             = _mm_add_ps(fjz1,tz);
 434
 435             /**************************
 436              * CALCULATE INTERACTIONS *
 437              **************************/
 438
 439             /* COULOMB ELECTROSTATICS */
 440             velec            = _mm_mul_ps(qq12,rinv12);
 441             felec            = _mm_mul_ps(velec,rinvsq12);
 442
 443             /* Update potential sum for this i atom from the interaction with this j atom. */
 444             velecsum         = _mm_add_ps(velecsum,velec);
 445
 446             fscal            = felec;
 447
 448             /* Calculate temporary vectorial force */
 449             tx               = _mm_mul_ps(fscal,dx12);
 450             ty               = _mm_mul_ps(fscal,dy12);
 451             tz               = _mm_mul_ps(fscal,dz12);
 452
 453             /* Update vectorial force */
 454             fix1             = _mm_add_ps(fix1,tx);
 455             fiy1             = _mm_add_ps(fiy1,ty);
 456             fiz1             = _mm_add_ps(fiz1,tz);
 457
 458             fjx2             = _mm_add_ps(fjx2,tx);
 459             fjy2             = _mm_add_ps(fjy2,ty);
 460             fjz2             = _mm_add_ps(fjz2,tz);
 461
 462             /**************************
 463              * CALCULATE INTERACTIONS *
 464              **************************/
 465
 466             /* COULOMB ELECTROSTATICS */
 467             velec            = _mm_mul_ps(qq20,rinv20);
 468             felec            = _mm_mul_ps(velec,rinvsq20);
 469
 470             /* Update potential sum for this i atom from the interaction with this j atom. */
 471             velecsum         = _mm_add_ps(velecsum,velec);
 472
 473             fscal            = felec;
 474
 475             /* Calculate temporary vectorial force */
 476             tx               = _mm_mul_ps(fscal,dx20);
 477             ty               = _mm_mul_ps(fscal,dy20);
 478             tz               = _mm_mul_ps(fscal,dz20);
 479
 480             /* Update vectorial force */
 481             fix2             = _mm_add_ps(fix2,tx);
 482             fiy2             = _mm_add_ps(fiy2,ty);
 483             fiz2             = _mm_add_ps(fiz2,tz);
 484
 485             fjx0             = _mm_add_ps(fjx0,tx);
 486             fjy0             = _mm_add_ps(fjy0,ty);
 487             fjz0             = _mm_add_ps(fjz0,tz);
 488
 489             /**************************
 490              * CALCULATE INTERACTIONS *
 491              **************************/
 492
 493             /* COULOMB ELECTROSTATICS */
 494             velec            = _mm_mul_ps(qq21,rinv21);
 495             felec            = _mm_mul_ps(velec,rinvsq21);
 496
 497             /* Update potential sum for this i atom from the interaction with this j atom. */
 498             velecsum         = _mm_add_ps(velecsum,velec);
 499
 500             fscal            = felec;
 501
 502             /* Calculate temporary vectorial force */
 503             tx               = _mm_mul_ps(fscal,dx21);
 504             ty               = _mm_mul_ps(fscal,dy21);
 505             tz               = _mm_mul_ps(fscal,dz21);
 506
 507             /* Update vectorial force */
 508             fix2             = _mm_add_ps(fix2,tx);
 509             fiy2             = _mm_add_ps(fiy2,ty);
 510             fiz2             = _mm_add_ps(fiz2,tz);
 511
 512             fjx1             = _mm_add_ps(fjx1,tx);
 513             fjy1             = _mm_add_ps(fjy1,ty);
 514             fjz1             = _mm_add_ps(fjz1,tz);
 515
 516             /**************************
 517              * CALCULATE INTERACTIONS *
 518              **************************/
 519
 520             /* COULOMB ELECTROSTATICS */
 521             velec            = _mm_mul_ps(qq22,rinv22);
 522             felec            = _mm_mul_ps(velec,rinvsq22);
 523
 524             /* Update potential sum for this i atom from the interaction with this j atom. */
 525             velecsum         = _mm_add_ps(velecsum,velec);
 526
 527             fscal            = felec;
 528
 529             /* Calculate temporary vectorial force */
 530             tx               = _mm_mul_ps(fscal,dx22);
 531             ty               = _mm_mul_ps(fscal,dy22);
 532             tz               = _mm_mul_ps(fscal,dz22);
 533
 534             /* Update vectorial force */
 535             fix2             = _mm_add_ps(fix2,tx);
 536             fiy2             = _mm_add_ps(fiy2,ty);
 537             fiz2             = _mm_add_ps(fiz2,tz);
 538
 539             fjx2             = _mm_add_ps(fjx2,tx);
 540             fjy2             = _mm_add_ps(fjy2,ty);
 541             fjz2             = _mm_add_ps(fjz2,tz);
 542
 543             fjptrA             = f+j_coord_offsetA;
 544             fjptrB             = f+j_coord_offsetB;
 545             fjptrC             = f+j_coord_offsetC;
 546             fjptrD             = f+j_coord_offsetD;
 547
 548             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 549                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 550
 551             /* Inner loop uses 264 flops */
 552         }
 553
 554         if(jidx<j_index_end)
 555         {
 556
 557             /* Get j neighbor index, and coordinate index */
 558             jnrlistA         = jjnr[jidx];
 559             jnrlistB         = jjnr[jidx+1];
 560             jnrlistC         = jjnr[jidx+2];
 561             jnrlistD         = jjnr[jidx+3];
 562             /* Sign of each element will be negative for non-real atoms.
 563              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 564              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 565              */
 566             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 567             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 568             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 569             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 570             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 571             j_coord_offsetA  = DIM*jnrA;
 572             j_coord_offsetB  = DIM*jnrB;
 573             j_coord_offsetC  = DIM*jnrC;
 574             j_coord_offsetD  = DIM*jnrD;
 575
 576             /* load j atom coordinates */
 577             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 578                                               x+j_coord_offsetC,x+j_coord_offsetD,
 579                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 580
 581             /* Calculate displacement vector */
 582             dx00             = _mm_sub_ps(ix0,jx0);
 583             dy00             = _mm_sub_ps(iy0,jy0);
 584             dz00             = _mm_sub_ps(iz0,jz0);
 585             dx01             = _mm_sub_ps(ix0,jx1);
 586             dy01             = _mm_sub_ps(iy0,jy1);
 587             dz01             = _mm_sub_ps(iz0,jz1);
 588             dx02             = _mm_sub_ps(ix0,jx2);
 589             dy02             = _mm_sub_ps(iy0,jy2);
 590             dz02             = _mm_sub_ps(iz0,jz2);
 591             dx10             = _mm_sub_ps(ix1,jx0);
 592             dy10             = _mm_sub_ps(iy1,jy0);
 593             dz10             = _mm_sub_ps(iz1,jz0);
 594             dx11             = _mm_sub_ps(ix1,jx1);
 595             dy11             = _mm_sub_ps(iy1,jy1);
 596             dz11             = _mm_sub_ps(iz1,jz1);
 597             dx12             = _mm_sub_ps(ix1,jx2);
 598             dy12             = _mm_sub_ps(iy1,jy2);
 599             dz12             = _mm_sub_ps(iz1,jz2);
 600             dx20             = _mm_sub_ps(ix2,jx0);
 601             dy20             = _mm_sub_ps(iy2,jy0);
 602             dz20             = _mm_sub_ps(iz2,jz0);
 603             dx21             = _mm_sub_ps(ix2,jx1);
 604             dy21             = _mm_sub_ps(iy2,jy1);
 605             dz21             = _mm_sub_ps(iz2,jz1);
 606             dx22             = _mm_sub_ps(ix2,jx2);
 607             dy22             = _mm_sub_ps(iy2,jy2);
 608             dz22             = _mm_sub_ps(iz2,jz2);
 609
 610             /* Calculate squared distance and things based on it */
 611             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 612             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 613             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 614             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 615             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 616             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 617             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 618             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 619             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 620
 621             rinv00           = sse41_invsqrt_f(rsq00);
 622             rinv01           = sse41_invsqrt_f(rsq01);
 623             rinv02           = sse41_invsqrt_f(rsq02);
 624             rinv10           = sse41_invsqrt_f(rsq10);
 625             rinv11           = sse41_invsqrt_f(rsq11);
 626             rinv12           = sse41_invsqrt_f(rsq12);
 627             rinv20           = sse41_invsqrt_f(rsq20);
 628             rinv21           = sse41_invsqrt_f(rsq21);
 629             rinv22           = sse41_invsqrt_f(rsq22);
 630
 631             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 632             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 633             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 634             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 635             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 636             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 637             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 638             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 639             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 640
 641             fjx0             = _mm_setzero_ps();
 642             fjy0             = _mm_setzero_ps();
 643             fjz0             = _mm_setzero_ps();
 644             fjx1             = _mm_setzero_ps();
 645             fjy1             = _mm_setzero_ps();
 646             fjz1             = _mm_setzero_ps();
 647             fjx2             = _mm_setzero_ps();
 648             fjy2             = _mm_setzero_ps();
 649             fjz2             = _mm_setzero_ps();
 650
 651             /**************************
 652              * CALCULATE INTERACTIONS *
 653              **************************/
 654
 655             /* COULOMB ELECTROSTATICS */
 656             velec            = _mm_mul_ps(qq00,rinv00);
 657             felec            = _mm_mul_ps(velec,rinvsq00);
 658
 659             /* LENNARD-JONES DISPERSION/REPULSION */
 660
 661             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 662             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 663             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 664             vvdw             = _mm_sub_ps( _mm_mul_ps(vvdw12,one_twelfth) , _mm_mul_ps(vvdw6,one_sixth) );
 665             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 666
 667             /* Update potential sum for this i atom from the interaction with this j atom. */
 668             velec            = _mm_andnot_ps(dummy_mask,velec);
 669             velecsum         = _mm_add_ps(velecsum,velec);
 670             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 671             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 672
 673             fscal            = _mm_add_ps(felec,fvdw);
 674
 675             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 676
 677             /* Calculate temporary vectorial force */
 678             tx               = _mm_mul_ps(fscal,dx00);
 679             ty               = _mm_mul_ps(fscal,dy00);
 680             tz               = _mm_mul_ps(fscal,dz00);
 681
 682             /* Update vectorial force */
 683             fix0             = _mm_add_ps(fix0,tx);
 684             fiy0             = _mm_add_ps(fiy0,ty);
 685             fiz0             = _mm_add_ps(fiz0,tz);
 686
 687             fjx0             = _mm_add_ps(fjx0,tx);
 688             fjy0             = _mm_add_ps(fjy0,ty);
 689             fjz0             = _mm_add_ps(fjz0,tz);
 690
 691             /**************************
 692              * CALCULATE INTERACTIONS *
 693              **************************/
 694
 695             /* COULOMB ELECTROSTATICS */
 696             velec            = _mm_mul_ps(qq01,rinv01);
 697             felec            = _mm_mul_ps(velec,rinvsq01);
 698
 699             /* Update potential sum for this i atom from the interaction with this j atom. */
 700             velec            = _mm_andnot_ps(dummy_mask,velec);
 701             velecsum         = _mm_add_ps(velecsum,velec);
 702
 703             fscal            = felec;
 704
 705             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 706
 707             /* Calculate temporary vectorial force */
 708             tx               = _mm_mul_ps(fscal,dx01);
 709             ty               = _mm_mul_ps(fscal,dy01);
 710             tz               = _mm_mul_ps(fscal,dz01);
 711
 712             /* Update vectorial force */
 713             fix0             = _mm_add_ps(fix0,tx);
 714             fiy0             = _mm_add_ps(fiy0,ty);
 715             fiz0             = _mm_add_ps(fiz0,tz);
 716
 717             fjx1             = _mm_add_ps(fjx1,tx);
 718             fjy1             = _mm_add_ps(fjy1,ty);
 719             fjz1             = _mm_add_ps(fjz1,tz);
 720
 721             /**************************
 722              * CALCULATE INTERACTIONS *
 723              **************************/
 724
 725             /* COULOMB ELECTROSTATICS */
 726             velec            = _mm_mul_ps(qq02,rinv02);
 727             felec            = _mm_mul_ps(velec,rinvsq02);
 728
 729             /* Update potential sum for this i atom from the interaction with this j atom. */
 730             velec            = _mm_andnot_ps(dummy_mask,velec);
 731             velecsum         = _mm_add_ps(velecsum,velec);
 732
 733             fscal            = felec;
 734
 735             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 736
 737             /* Calculate temporary vectorial force */
 738             tx               = _mm_mul_ps(fscal,dx02);
 739             ty               = _mm_mul_ps(fscal,dy02);
 740             tz               = _mm_mul_ps(fscal,dz02);
 741
 742             /* Update vectorial force */
 743             fix0             = _mm_add_ps(fix0,tx);
 744             fiy0             = _mm_add_ps(fiy0,ty);
 745             fiz0             = _mm_add_ps(fiz0,tz);
 746
 747             fjx2             = _mm_add_ps(fjx2,tx);
 748             fjy2             = _mm_add_ps(fjy2,ty);
 749             fjz2             = _mm_add_ps(fjz2,tz);
 750
 751             /**************************
 752              * CALCULATE INTERACTIONS *
 753              **************************/
 754
 755             /* COULOMB ELECTROSTATICS */
 756             velec            = _mm_mul_ps(qq10,rinv10);
 757             felec            = _mm_mul_ps(velec,rinvsq10);
 758
 759             /* Update potential sum for this i atom from the interaction with this j atom. */
 760             velec            = _mm_andnot_ps(dummy_mask,velec);
 761             velecsum         = _mm_add_ps(velecsum,velec);
 762
 763             fscal            = felec;
 764
 765             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 766
 767             /* Calculate temporary vectorial force */
 768             tx               = _mm_mul_ps(fscal,dx10);
 769             ty               = _mm_mul_ps(fscal,dy10);
 770             tz               = _mm_mul_ps(fscal,dz10);
 771
 772             /* Update vectorial force */
 773             fix1             = _mm_add_ps(fix1,tx);
 774             fiy1             = _mm_add_ps(fiy1,ty);
 775             fiz1             = _mm_add_ps(fiz1,tz);
 776
 777             fjx0             = _mm_add_ps(fjx0,tx);
 778             fjy0             = _mm_add_ps(fjy0,ty);
 779             fjz0             = _mm_add_ps(fjz0,tz);
 780
 781             /**************************
 782              * CALCULATE INTERACTIONS *
 783              **************************/
 784
 785             /* COULOMB ELECTROSTATICS */
 786             velec            = _mm_mul_ps(qq11,rinv11);
 787             felec            = _mm_mul_ps(velec,rinvsq11);
 788
 789             /* Update potential sum for this i atom from the interaction with this j atom. */
 790             velec            = _mm_andnot_ps(dummy_mask,velec);
 791             velecsum         = _mm_add_ps(velecsum,velec);
 792
 793             fscal            = felec;
 794
 795             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 796
 797             /* Calculate temporary vectorial force */
 798             tx               = _mm_mul_ps(fscal,dx11);
 799             ty               = _mm_mul_ps(fscal,dy11);
 800             tz               = _mm_mul_ps(fscal,dz11);
 801
 802             /* Update vectorial force */
 803             fix1             = _mm_add_ps(fix1,tx);
 804             fiy1             = _mm_add_ps(fiy1,ty);
 805             fiz1             = _mm_add_ps(fiz1,tz);
 806
 807             fjx1             = _mm_add_ps(fjx1,tx);
 808             fjy1             = _mm_add_ps(fjy1,ty);
 809             fjz1             = _mm_add_ps(fjz1,tz);
 810
 811             /**************************
 812              * CALCULATE INTERACTIONS *
 813              **************************/
 814
 815             /* COULOMB ELECTROSTATICS */
 816             velec            = _mm_mul_ps(qq12,rinv12);
 817             felec            = _mm_mul_ps(velec,rinvsq12);
 818
 819             /* Update potential sum for this i atom from the interaction with this j atom. */
 820             velec            = _mm_andnot_ps(dummy_mask,velec);
 821             velecsum         = _mm_add_ps(velecsum,velec);
 822
 823             fscal            = felec;
 824
 825             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 826
 827             /* Calculate temporary vectorial force */
 828             tx               = _mm_mul_ps(fscal,dx12);
 829             ty               = _mm_mul_ps(fscal,dy12);
 830             tz               = _mm_mul_ps(fscal,dz12);
 831
 832             /* Update vectorial force */
 833             fix1             = _mm_add_ps(fix1,tx);
 834             fiy1             = _mm_add_ps(fiy1,ty);
 835             fiz1             = _mm_add_ps(fiz1,tz);
 836
 837             fjx2             = _mm_add_ps(fjx2,tx);
 838             fjy2             = _mm_add_ps(fjy2,ty);
 839             fjz2             = _mm_add_ps(fjz2,tz);
 840
 841             /**************************
 842              * CALCULATE INTERACTIONS *
 843              **************************/
 844
 845             /* COULOMB ELECTROSTATICS */
 846             velec            = _mm_mul_ps(qq20,rinv20);
 847             felec            = _mm_mul_ps(velec,rinvsq20);
 848
 849             /* Update potential sum for this i atom from the interaction with this j atom. */
 850             velec            = _mm_andnot_ps(dummy_mask,velec);
 851             velecsum         = _mm_add_ps(velecsum,velec);
 852
 853             fscal            = felec;
 854
 855             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 856
 857             /* Calculate temporary vectorial force */
 858             tx               = _mm_mul_ps(fscal,dx20);
 859             ty               = _mm_mul_ps(fscal,dy20);
 860             tz               = _mm_mul_ps(fscal,dz20);
 861
 862             /* Update vectorial force */
 863             fix2             = _mm_add_ps(fix2,tx);
 864             fiy2             = _mm_add_ps(fiy2,ty);
 865             fiz2             = _mm_add_ps(fiz2,tz);
 866
 867             fjx0             = _mm_add_ps(fjx0,tx);
 868             fjy0             = _mm_add_ps(fjy0,ty);
 869             fjz0             = _mm_add_ps(fjz0,tz);
 870
 871             /**************************
 872              * CALCULATE INTERACTIONS *
 873              **************************/
 874
 875             /* COULOMB ELECTROSTATICS */
 876             velec            = _mm_mul_ps(qq21,rinv21);
 877             felec            = _mm_mul_ps(velec,rinvsq21);
 878
 879             /* Update potential sum for this i atom from the interaction with this j atom. */
 880             velec            = _mm_andnot_ps(dummy_mask,velec);
 881             velecsum         = _mm_add_ps(velecsum,velec);
 882
 883             fscal            = felec;
 884
 885             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 886
 887             /* Calculate temporary vectorial force */
 888             tx               = _mm_mul_ps(fscal,dx21);
 889             ty               = _mm_mul_ps(fscal,dy21);
 890             tz               = _mm_mul_ps(fscal,dz21);
 891
 892             /* Update vectorial force */
 893             fix2             = _mm_add_ps(fix2,tx);
 894             fiy2             = _mm_add_ps(fiy2,ty);
 895             fiz2             = _mm_add_ps(fiz2,tz);
 896
 897             fjx1             = _mm_add_ps(fjx1,tx);
 898             fjy1             = _mm_add_ps(fjy1,ty);
 899             fjz1             = _mm_add_ps(fjz1,tz);
 900
 901             /**************************
 902              * CALCULATE INTERACTIONS *
 903              **************************/
 904
 905             /* COULOMB ELECTROSTATICS */
 906             velec            = _mm_mul_ps(qq22,rinv22);
 907             felec            = _mm_mul_ps(velec,rinvsq22);
 908
 909             /* Update potential sum for this i atom from the interaction with this j atom. */
 910             velec            = _mm_andnot_ps(dummy_mask,velec);
 911             velecsum         = _mm_add_ps(velecsum,velec);
 912
 913             fscal            = felec;
 914
 915             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 916
 917             /* Calculate temporary vectorial force */
 918             tx               = _mm_mul_ps(fscal,dx22);
 919             ty               = _mm_mul_ps(fscal,dy22);
 920             tz               = _mm_mul_ps(fscal,dz22);
 921
 922             /* Update vectorial force */
 923             fix2             = _mm_add_ps(fix2,tx);
 924             fiy2             = _mm_add_ps(fiy2,ty);
 925             fiz2             = _mm_add_ps(fiz2,tz);
 926
 927             fjx2             = _mm_add_ps(fjx2,tx);
 928             fjy2             = _mm_add_ps(fjy2,ty);
 929             fjz2             = _mm_add_ps(fjz2,tz);
 930
 931             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 932             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 933             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 934             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 935
 936             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 937                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 938
 939             /* Inner loop uses 264 flops */
 940         }
 941
 942         /* End of innermost loop */
 943
 944         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 945                                               f+i_coord_offset,fshift+i_shift_offset);
 946
 947         ggid                        = gid[iidx];
 948         /* Update potential energies */
 949         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 950         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 951
 952         /* Increment number of inner iterations */
 953         inneriter                  += j_index_end - j_index_start;
 954
 955         /* Outer loop uses 20 flops */
 956     }
 957
 958     /* Increment number of outer iterations */
 959     outeriter        += nri;
 960
 961     /* Update outer/inner flops */
 962
 963     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*264);
 964 }
 965 /*
 966  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW3W3_F_sse4_1_single
 967  * Electrostatics interaction: Coulomb
 968  * VdW interaction:            LennardJones
 969  * Geometry:                   Water3-Water3
 970  * Calculate force/pot:        Force
 971  */
 972 void
 973 nb_kernel_ElecCoul_VdwLJ_GeomW3W3_F_sse4_1_single
 974                     (t_nblist                    * gmx_restrict       nlist,
 975                      rvec                        * gmx_restrict          xx,
 976                      rvec                        * gmx_restrict          ff,
 977                      struct t_forcerec           * gmx_restrict          fr,
 978                      t_mdatoms                   * gmx_restrict     mdatoms,
 979                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 980                      t_nrnb                      * gmx_restrict        nrnb)
 981 {
 982     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 983      * just 0 for non-waters.
 984      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 985      * jnr indices corresponding to data put in the four positions in the SIMD register.
 986      */
 987     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 988     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 989     int              jnrA,jnrB,jnrC,jnrD;
 990     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 991     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 992     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 993     real             rcutoff_scalar;
 994     real             *shiftvec,*fshift,*x,*f;
 995     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 996     real             scratch[4*DIM];
 997     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 998     int              vdwioffset0;
 999     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1000     int              vdwioffset1;
1001     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1002     int              vdwioffset2;
1003     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1004     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1005     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1006     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1007     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1008     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1009     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1010     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1011     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1012     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1013     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1014     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1015     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1016     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1017     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1018     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1019     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1020     real             *charge;
1021     int              nvdwtype;
1022     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1023     int              *vdwtype;
1024     real             *vdwparam;
1025     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1026     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1027     __m128           dummy_mask,cutoff_mask;
1028     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1029     __m128           one     = _mm_set1_ps(1.0);
1030     __m128           two     = _mm_set1_ps(2.0);
1031     x                = xx[0];
1032     f                = ff[0];
1033
1034     nri              = nlist->nri;
1035     iinr             = nlist->iinr;
1036     jindex           = nlist->jindex;
1037     jjnr             = nlist->jjnr;
1038     shiftidx         = nlist->shift;
1039     gid              = nlist->gid;
1040     shiftvec         = fr->shift_vec[0];
1041     fshift           = fr->fshift[0];
1042     facel            = _mm_set1_ps(fr->ic->epsfac);
1043     charge           = mdatoms->chargeA;
1044     nvdwtype         = fr->ntype;
1045     vdwparam         = fr->nbfp;
1046     vdwtype          = mdatoms->typeA;
1047
1048     /* Setup water-specific parameters */
1049     inr              = nlist->iinr[0];
1050     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1051     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1052     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1053     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1054
1055     jq0              = _mm_set1_ps(charge[inr+0]);
1056     jq1              = _mm_set1_ps(charge[inr+1]);
1057     jq2              = _mm_set1_ps(charge[inr+2]);
1058     vdwjidx0A        = 2*vdwtype[inr+0];
1059     qq00             = _mm_mul_ps(iq0,jq0);
1060     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1061     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1062     qq01             = _mm_mul_ps(iq0,jq1);
1063     qq02             = _mm_mul_ps(iq0,jq2);
1064     qq10             = _mm_mul_ps(iq1,jq0);
1065     qq11             = _mm_mul_ps(iq1,jq1);
1066     qq12             = _mm_mul_ps(iq1,jq2);
1067     qq20             = _mm_mul_ps(iq2,jq0);
1068     qq21             = _mm_mul_ps(iq2,jq1);
1069     qq22             = _mm_mul_ps(iq2,jq2);
1070
1071     /* Avoid stupid compiler warnings */
1072     jnrA = jnrB = jnrC = jnrD = 0;
1073     j_coord_offsetA = 0;
1074     j_coord_offsetB = 0;
1075     j_coord_offsetC = 0;
1076     j_coord_offsetD = 0;
1077
1078     outeriter        = 0;
1079     inneriter        = 0;
1080
1081     for(iidx=0;iidx<4*DIM;iidx++)
1082     {
1083         scratch[iidx] = 0.0;
1084     }
1085
1086     /* Start outer loop over neighborlists */
1087     for(iidx=0; iidx<nri; iidx++)
1088     {
1089         /* Load shift vector for this list */
1090         i_shift_offset   = DIM*shiftidx[iidx];
1091
1092         /* Load limits for loop over neighbors */
1093         j_index_start    = jindex[iidx];
1094         j_index_end      = jindex[iidx+1];
1095
1096         /* Get outer coordinate index */
1097         inr              = iinr[iidx];
1098         i_coord_offset   = DIM*inr;
1099
1100         /* Load i particle coords and add shift vector */
1101         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1102                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1103
1104         fix0             = _mm_setzero_ps();
1105         fiy0             = _mm_setzero_ps();
1106         fiz0             = _mm_setzero_ps();
1107         fix1             = _mm_setzero_ps();
1108         fiy1             = _mm_setzero_ps();
1109         fiz1             = _mm_setzero_ps();
1110         fix2             = _mm_setzero_ps();
1111         fiy2             = _mm_setzero_ps();
1112         fiz2             = _mm_setzero_ps();
1113
1114         /* Start inner kernel loop */
1115         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1116         {
1117
1118             /* Get j neighbor index, and coordinate index */
1119             jnrA             = jjnr[jidx];
1120             jnrB             = jjnr[jidx+1];
1121             jnrC             = jjnr[jidx+2];
1122             jnrD             = jjnr[jidx+3];
1123             j_coord_offsetA  = DIM*jnrA;
1124             j_coord_offsetB  = DIM*jnrB;
1125             j_coord_offsetC  = DIM*jnrC;
1126             j_coord_offsetD  = DIM*jnrD;
1127
1128             /* load j atom coordinates */
1129             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1130                                               x+j_coord_offsetC,x+j_coord_offsetD,
1131                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1132
1133             /* Calculate displacement vector */
1134             dx00             = _mm_sub_ps(ix0,jx0);
1135             dy00             = _mm_sub_ps(iy0,jy0);
1136             dz00             = _mm_sub_ps(iz0,jz0);
1137             dx01             = _mm_sub_ps(ix0,jx1);
1138             dy01             = _mm_sub_ps(iy0,jy1);
1139             dz01             = _mm_sub_ps(iz0,jz1);
1140             dx02             = _mm_sub_ps(ix0,jx2);
1141             dy02             = _mm_sub_ps(iy0,jy2);
1142             dz02             = _mm_sub_ps(iz0,jz2);
1143             dx10             = _mm_sub_ps(ix1,jx0);
1144             dy10             = _mm_sub_ps(iy1,jy0);
1145             dz10             = _mm_sub_ps(iz1,jz0);
1146             dx11             = _mm_sub_ps(ix1,jx1);
1147             dy11             = _mm_sub_ps(iy1,jy1);
1148             dz11             = _mm_sub_ps(iz1,jz1);
1149             dx12             = _mm_sub_ps(ix1,jx2);
1150             dy12             = _mm_sub_ps(iy1,jy2);
1151             dz12             = _mm_sub_ps(iz1,jz2);
1152             dx20             = _mm_sub_ps(ix2,jx0);
1153             dy20             = _mm_sub_ps(iy2,jy0);
1154             dz20             = _mm_sub_ps(iz2,jz0);
1155             dx21             = _mm_sub_ps(ix2,jx1);
1156             dy21             = _mm_sub_ps(iy2,jy1);
1157             dz21             = _mm_sub_ps(iz2,jz1);
1158             dx22             = _mm_sub_ps(ix2,jx2);
1159             dy22             = _mm_sub_ps(iy2,jy2);
1160             dz22             = _mm_sub_ps(iz2,jz2);
1161
1162             /* Calculate squared distance and things based on it */
1163             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1164             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1165             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1166             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1167             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1168             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1169             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1170             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1171             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1172
1173             rinv00           = sse41_invsqrt_f(rsq00);
1174             rinv01           = sse41_invsqrt_f(rsq01);
1175             rinv02           = sse41_invsqrt_f(rsq02);
1176             rinv10           = sse41_invsqrt_f(rsq10);
1177             rinv11           = sse41_invsqrt_f(rsq11);
1178             rinv12           = sse41_invsqrt_f(rsq12);
1179             rinv20           = sse41_invsqrt_f(rsq20);
1180             rinv21           = sse41_invsqrt_f(rsq21);
1181             rinv22           = sse41_invsqrt_f(rsq22);
1182
1183             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1184             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1185             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1186             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1187             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1188             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1189             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1190             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1191             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1192
1193             fjx0             = _mm_setzero_ps();
1194             fjy0             = _mm_setzero_ps();
1195             fjz0             = _mm_setzero_ps();
1196             fjx1             = _mm_setzero_ps();
1197             fjy1             = _mm_setzero_ps();
1198             fjz1             = _mm_setzero_ps();
1199             fjx2             = _mm_setzero_ps();
1200             fjy2             = _mm_setzero_ps();
1201             fjz2             = _mm_setzero_ps();
1202
1203             /**************************
1204              * CALCULATE INTERACTIONS *
1205              **************************/
1206
1207             /* COULOMB ELECTROSTATICS */
1208             velec            = _mm_mul_ps(qq00,rinv00);
1209             felec            = _mm_mul_ps(velec,rinvsq00);
1210
1211             /* LENNARD-JONES DISPERSION/REPULSION */
1212
1213             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1214             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1215
1216             fscal            = _mm_add_ps(felec,fvdw);
1217
1218             /* Calculate temporary vectorial force */
1219             tx               = _mm_mul_ps(fscal,dx00);
1220             ty               = _mm_mul_ps(fscal,dy00);
1221             tz               = _mm_mul_ps(fscal,dz00);
1222
1223             /* Update vectorial force */
1224             fix0             = _mm_add_ps(fix0,tx);
1225             fiy0             = _mm_add_ps(fiy0,ty);
1226             fiz0             = _mm_add_ps(fiz0,tz);
1227
1228             fjx0             = _mm_add_ps(fjx0,tx);
1229             fjy0             = _mm_add_ps(fjy0,ty);
1230             fjz0             = _mm_add_ps(fjz0,tz);
1231
1232             /**************************
1233              * CALCULATE INTERACTIONS *
1234              **************************/
1235
1236             /* COULOMB ELECTROSTATICS */
1237             velec            = _mm_mul_ps(qq01,rinv01);
1238             felec            = _mm_mul_ps(velec,rinvsq01);
1239
1240             fscal            = felec;
1241
1242             /* Calculate temporary vectorial force */
1243             tx               = _mm_mul_ps(fscal,dx01);
1244             ty               = _mm_mul_ps(fscal,dy01);
1245             tz               = _mm_mul_ps(fscal,dz01);
1246
1247             /* Update vectorial force */
1248             fix0             = _mm_add_ps(fix0,tx);
1249             fiy0             = _mm_add_ps(fiy0,ty);
1250             fiz0             = _mm_add_ps(fiz0,tz);
1251
1252             fjx1             = _mm_add_ps(fjx1,tx);
1253             fjy1             = _mm_add_ps(fjy1,ty);
1254             fjz1             = _mm_add_ps(fjz1,tz);
1255
1256             /**************************
1257              * CALCULATE INTERACTIONS *
1258              **************************/
1259
1260             /* COULOMB ELECTROSTATICS */
1261             velec            = _mm_mul_ps(qq02,rinv02);
1262             felec            = _mm_mul_ps(velec,rinvsq02);
1263
1264             fscal            = felec;
1265
1266             /* Calculate temporary vectorial force */
1267             tx               = _mm_mul_ps(fscal,dx02);
1268             ty               = _mm_mul_ps(fscal,dy02);
1269             tz               = _mm_mul_ps(fscal,dz02);
1270
1271             /* Update vectorial force */
1272             fix0             = _mm_add_ps(fix0,tx);
1273             fiy0             = _mm_add_ps(fiy0,ty);
1274             fiz0             = _mm_add_ps(fiz0,tz);
1275
1276             fjx2             = _mm_add_ps(fjx2,tx);
1277             fjy2             = _mm_add_ps(fjy2,ty);
1278             fjz2             = _mm_add_ps(fjz2,tz);
1279
1280             /**************************
1281              * CALCULATE INTERACTIONS *
1282              **************************/
1283
1284             /* COULOMB ELECTROSTATICS */
1285             velec            = _mm_mul_ps(qq10,rinv10);
1286             felec            = _mm_mul_ps(velec,rinvsq10);
1287
1288             fscal            = felec;
1289
1290             /* Calculate temporary vectorial force */
1291             tx               = _mm_mul_ps(fscal,dx10);
1292             ty               = _mm_mul_ps(fscal,dy10);
1293             tz               = _mm_mul_ps(fscal,dz10);
1294
1295             /* Update vectorial force */
1296             fix1             = _mm_add_ps(fix1,tx);
1297             fiy1             = _mm_add_ps(fiy1,ty);
1298             fiz1             = _mm_add_ps(fiz1,tz);
1299
1300             fjx0             = _mm_add_ps(fjx0,tx);
1301             fjy0             = _mm_add_ps(fjy0,ty);
1302             fjz0             = _mm_add_ps(fjz0,tz);
1303
1304             /**************************
1305              * CALCULATE INTERACTIONS *
1306              **************************/
1307
1308             /* COULOMB ELECTROSTATICS */
1309             velec            = _mm_mul_ps(qq11,rinv11);
1310             felec            = _mm_mul_ps(velec,rinvsq11);
1311
1312             fscal            = felec;
1313
1314             /* Calculate temporary vectorial force */
1315             tx               = _mm_mul_ps(fscal,dx11);
1316             ty               = _mm_mul_ps(fscal,dy11);
1317             tz               = _mm_mul_ps(fscal,dz11);
1318
1319             /* Update vectorial force */
1320             fix1             = _mm_add_ps(fix1,tx);
1321             fiy1             = _mm_add_ps(fiy1,ty);
1322             fiz1             = _mm_add_ps(fiz1,tz);
1323
1324             fjx1             = _mm_add_ps(fjx1,tx);
1325             fjy1             = _mm_add_ps(fjy1,ty);
1326             fjz1             = _mm_add_ps(fjz1,tz);
1327
1328             /**************************
1329              * CALCULATE INTERACTIONS *
1330              **************************/
1331
1332             /* COULOMB ELECTROSTATICS */
1333             velec            = _mm_mul_ps(qq12,rinv12);
1334             felec            = _mm_mul_ps(velec,rinvsq12);
1335
1336             fscal            = felec;
1337
1338             /* Calculate temporary vectorial force */
1339             tx               = _mm_mul_ps(fscal,dx12);
1340             ty               = _mm_mul_ps(fscal,dy12);
1341             tz               = _mm_mul_ps(fscal,dz12);
1342
1343             /* Update vectorial force */
1344             fix1             = _mm_add_ps(fix1,tx);
1345             fiy1             = _mm_add_ps(fiy1,ty);
1346             fiz1             = _mm_add_ps(fiz1,tz);
1347
1348             fjx2             = _mm_add_ps(fjx2,tx);
1349             fjy2             = _mm_add_ps(fjy2,ty);
1350             fjz2             = _mm_add_ps(fjz2,tz);
1351
1352             /**************************
1353              * CALCULATE INTERACTIONS *
1354              **************************/
1355
1356             /* COULOMB ELECTROSTATICS */
1357             velec            = _mm_mul_ps(qq20,rinv20);
1358             felec            = _mm_mul_ps(velec,rinvsq20);
1359
1360             fscal            = felec;
1361
1362             /* Calculate temporary vectorial force */
1363             tx               = _mm_mul_ps(fscal,dx20);
1364             ty               = _mm_mul_ps(fscal,dy20);
1365             tz               = _mm_mul_ps(fscal,dz20);
1366
1367             /* Update vectorial force */
1368             fix2             = _mm_add_ps(fix2,tx);
1369             fiy2             = _mm_add_ps(fiy2,ty);
1370             fiz2             = _mm_add_ps(fiz2,tz);
1371
1372             fjx0             = _mm_add_ps(fjx0,tx);
1373             fjy0             = _mm_add_ps(fjy0,ty);
1374             fjz0             = _mm_add_ps(fjz0,tz);
1375
1376             /**************************
1377              * CALCULATE INTERACTIONS *
1378              **************************/
1379
1380             /* COULOMB ELECTROSTATICS */
1381             velec            = _mm_mul_ps(qq21,rinv21);
1382             felec            = _mm_mul_ps(velec,rinvsq21);
1383
1384             fscal            = felec;
1385
1386             /* Calculate temporary vectorial force */
1387             tx               = _mm_mul_ps(fscal,dx21);
1388             ty               = _mm_mul_ps(fscal,dy21);
1389             tz               = _mm_mul_ps(fscal,dz21);
1390
1391             /* Update vectorial force */
1392             fix2             = _mm_add_ps(fix2,tx);
1393             fiy2             = _mm_add_ps(fiy2,ty);
1394             fiz2             = _mm_add_ps(fiz2,tz);
1395
1396             fjx1             = _mm_add_ps(fjx1,tx);
1397             fjy1             = _mm_add_ps(fjy1,ty);
1398             fjz1             = _mm_add_ps(fjz1,tz);
1399
1400             /**************************
1401              * CALCULATE INTERACTIONS *
1402              **************************/
1403
1404             /* COULOMB ELECTROSTATICS */
1405             velec            = _mm_mul_ps(qq22,rinv22);
1406             felec            = _mm_mul_ps(velec,rinvsq22);
1407
1408             fscal            = felec;
1409
1410             /* Calculate temporary vectorial force */
1411             tx               = _mm_mul_ps(fscal,dx22);
1412             ty               = _mm_mul_ps(fscal,dy22);
1413             tz               = _mm_mul_ps(fscal,dz22);
1414
1415             /* Update vectorial force */
1416             fix2             = _mm_add_ps(fix2,tx);
1417             fiy2             = _mm_add_ps(fiy2,ty);
1418             fiz2             = _mm_add_ps(fiz2,tz);
1419
1420             fjx2             = _mm_add_ps(fjx2,tx);
1421             fjy2             = _mm_add_ps(fjy2,ty);
1422             fjz2             = _mm_add_ps(fjz2,tz);
1423
1424             fjptrA             = f+j_coord_offsetA;
1425             fjptrB             = f+j_coord_offsetB;
1426             fjptrC             = f+j_coord_offsetC;
1427             fjptrD             = f+j_coord_offsetD;
1428
1429             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1430                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1431
1432             /* Inner loop uses 250 flops */
1433         }
1434
1435         if(jidx<j_index_end)
1436         {
1437
1438             /* Get j neighbor index, and coordinate index */
1439             jnrlistA         = jjnr[jidx];
1440             jnrlistB         = jjnr[jidx+1];
1441             jnrlistC         = jjnr[jidx+2];
1442             jnrlistD         = jjnr[jidx+3];
1443             /* Sign of each element will be negative for non-real atoms.
1444              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1445              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1446              */
1447             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1448             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1449             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1450             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1451             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1452             j_coord_offsetA  = DIM*jnrA;
1453             j_coord_offsetB  = DIM*jnrB;
1454             j_coord_offsetC  = DIM*jnrC;
1455             j_coord_offsetD  = DIM*jnrD;
1456
1457             /* load j atom coordinates */
1458             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1459                                               x+j_coord_offsetC,x+j_coord_offsetD,
1460                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1461
1462             /* Calculate displacement vector */
1463             dx00             = _mm_sub_ps(ix0,jx0);
1464             dy00             = _mm_sub_ps(iy0,jy0);
1465             dz00             = _mm_sub_ps(iz0,jz0);
1466             dx01             = _mm_sub_ps(ix0,jx1);
1467             dy01             = _mm_sub_ps(iy0,jy1);
1468             dz01             = _mm_sub_ps(iz0,jz1);
1469             dx02             = _mm_sub_ps(ix0,jx2);
1470             dy02             = _mm_sub_ps(iy0,jy2);
1471             dz02             = _mm_sub_ps(iz0,jz2);
1472             dx10             = _mm_sub_ps(ix1,jx0);
1473             dy10             = _mm_sub_ps(iy1,jy0);
1474             dz10             = _mm_sub_ps(iz1,jz0);
1475             dx11             = _mm_sub_ps(ix1,jx1);
1476             dy11             = _mm_sub_ps(iy1,jy1);
1477             dz11             = _mm_sub_ps(iz1,jz1);
1478             dx12             = _mm_sub_ps(ix1,jx2);
1479             dy12             = _mm_sub_ps(iy1,jy2);
1480             dz12             = _mm_sub_ps(iz1,jz2);
1481             dx20             = _mm_sub_ps(ix2,jx0);
1482             dy20             = _mm_sub_ps(iy2,jy0);
1483             dz20             = _mm_sub_ps(iz2,jz0);
1484             dx21             = _mm_sub_ps(ix2,jx1);
1485             dy21             = _mm_sub_ps(iy2,jy1);
1486             dz21             = _mm_sub_ps(iz2,jz1);
1487             dx22             = _mm_sub_ps(ix2,jx2);
1488             dy22             = _mm_sub_ps(iy2,jy2);
1489             dz22             = _mm_sub_ps(iz2,jz2);
1490
1491             /* Calculate squared distance and things based on it */
1492             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1493             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1494             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1495             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1496             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1497             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1498             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1499             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1500             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1501
1502             rinv00           = sse41_invsqrt_f(rsq00);
1503             rinv01           = sse41_invsqrt_f(rsq01);
1504             rinv02           = sse41_invsqrt_f(rsq02);
1505             rinv10           = sse41_invsqrt_f(rsq10);
1506             rinv11           = sse41_invsqrt_f(rsq11);
1507             rinv12           = sse41_invsqrt_f(rsq12);
1508             rinv20           = sse41_invsqrt_f(rsq20);
1509             rinv21           = sse41_invsqrt_f(rsq21);
1510             rinv22           = sse41_invsqrt_f(rsq22);
1511
1512             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1513             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1514             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1515             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1516             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1517             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1518             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1519             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1520             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1521
1522             fjx0             = _mm_setzero_ps();
1523             fjy0             = _mm_setzero_ps();
1524             fjz0             = _mm_setzero_ps();
1525             fjx1             = _mm_setzero_ps();
1526             fjy1             = _mm_setzero_ps();
1527             fjz1             = _mm_setzero_ps();
1528             fjx2             = _mm_setzero_ps();
1529             fjy2             = _mm_setzero_ps();
1530             fjz2             = _mm_setzero_ps();
1531
1532             /**************************
1533              * CALCULATE INTERACTIONS *
1534              **************************/
1535
1536             /* COULOMB ELECTROSTATICS */
1537             velec            = _mm_mul_ps(qq00,rinv00);
1538             felec            = _mm_mul_ps(velec,rinvsq00);
1539
1540             /* LENNARD-JONES DISPERSION/REPULSION */
1541
1542             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1543             fvdw             = _mm_mul_ps(_mm_sub_ps(_mm_mul_ps(c12_00,rinvsix),c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1544
1545             fscal            = _mm_add_ps(felec,fvdw);
1546
1547             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1548
1549             /* Calculate temporary vectorial force */
1550             tx               = _mm_mul_ps(fscal,dx00);
1551             ty               = _mm_mul_ps(fscal,dy00);
1552             tz               = _mm_mul_ps(fscal,dz00);
1553
1554             /* Update vectorial force */
1555             fix0             = _mm_add_ps(fix0,tx);
1556             fiy0             = _mm_add_ps(fiy0,ty);
1557             fiz0             = _mm_add_ps(fiz0,tz);
1558
1559             fjx0             = _mm_add_ps(fjx0,tx);
1560             fjy0             = _mm_add_ps(fjy0,ty);
1561             fjz0             = _mm_add_ps(fjz0,tz);
1562
1563             /**************************
1564              * CALCULATE INTERACTIONS *
1565              **************************/
1566
1567             /* COULOMB ELECTROSTATICS */
1568             velec            = _mm_mul_ps(qq01,rinv01);
1569             felec            = _mm_mul_ps(velec,rinvsq01);
1570
1571             fscal            = felec;
1572
1573             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1574
1575             /* Calculate temporary vectorial force */
1576             tx               = _mm_mul_ps(fscal,dx01);
1577             ty               = _mm_mul_ps(fscal,dy01);
1578             tz               = _mm_mul_ps(fscal,dz01);
1579
1580             /* Update vectorial force */
1581             fix0             = _mm_add_ps(fix0,tx);
1582             fiy0             = _mm_add_ps(fiy0,ty);
1583             fiz0             = _mm_add_ps(fiz0,tz);
1584
1585             fjx1             = _mm_add_ps(fjx1,tx);
1586             fjy1             = _mm_add_ps(fjy1,ty);
1587             fjz1             = _mm_add_ps(fjz1,tz);
1588
1589             /**************************
1590              * CALCULATE INTERACTIONS *
1591              **************************/
1592
1593             /* COULOMB ELECTROSTATICS */
1594             velec            = _mm_mul_ps(qq02,rinv02);
1595             felec            = _mm_mul_ps(velec,rinvsq02);
1596
1597             fscal            = felec;
1598
1599             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1600
1601             /* Calculate temporary vectorial force */
1602             tx               = _mm_mul_ps(fscal,dx02);
1603             ty               = _mm_mul_ps(fscal,dy02);
1604             tz               = _mm_mul_ps(fscal,dz02);
1605
1606             /* Update vectorial force */
1607             fix0             = _mm_add_ps(fix0,tx);
1608             fiy0             = _mm_add_ps(fiy0,ty);
1609             fiz0             = _mm_add_ps(fiz0,tz);
1610
1611             fjx2             = _mm_add_ps(fjx2,tx);
1612             fjy2             = _mm_add_ps(fjy2,ty);
1613             fjz2             = _mm_add_ps(fjz2,tz);
1614
1615             /**************************
1616              * CALCULATE INTERACTIONS *
1617              **************************/
1618
1619             /* COULOMB ELECTROSTATICS */
1620             velec            = _mm_mul_ps(qq10,rinv10);
1621             felec            = _mm_mul_ps(velec,rinvsq10);
1622
1623             fscal            = felec;
1624
1625             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1626
1627             /* Calculate temporary vectorial force */
1628             tx               = _mm_mul_ps(fscal,dx10);
1629             ty               = _mm_mul_ps(fscal,dy10);
1630             tz               = _mm_mul_ps(fscal,dz10);
1631
1632             /* Update vectorial force */
1633             fix1             = _mm_add_ps(fix1,tx);
1634             fiy1             = _mm_add_ps(fiy1,ty);
1635             fiz1             = _mm_add_ps(fiz1,tz);
1636
1637             fjx0             = _mm_add_ps(fjx0,tx);
1638             fjy0             = _mm_add_ps(fjy0,ty);
1639             fjz0             = _mm_add_ps(fjz0,tz);
1640
1641             /**************************
1642              * CALCULATE INTERACTIONS *
1643              **************************/
1644
1645             /* COULOMB ELECTROSTATICS */
1646             velec            = _mm_mul_ps(qq11,rinv11);
1647             felec            = _mm_mul_ps(velec,rinvsq11);
1648
1649             fscal            = felec;
1650
1651             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1652
1653             /* Calculate temporary vectorial force */
1654             tx               = _mm_mul_ps(fscal,dx11);
1655             ty               = _mm_mul_ps(fscal,dy11);
1656             tz               = _mm_mul_ps(fscal,dz11);
1657
1658             /* Update vectorial force */
1659             fix1             = _mm_add_ps(fix1,tx);
1660             fiy1             = _mm_add_ps(fiy1,ty);
1661             fiz1             = _mm_add_ps(fiz1,tz);
1662
1663             fjx1             = _mm_add_ps(fjx1,tx);
1664             fjy1             = _mm_add_ps(fjy1,ty);
1665             fjz1             = _mm_add_ps(fjz1,tz);
1666
1667             /**************************
1668              * CALCULATE INTERACTIONS *
1669              **************************/
1670
1671             /* COULOMB ELECTROSTATICS */
1672             velec            = _mm_mul_ps(qq12,rinv12);
1673             felec            = _mm_mul_ps(velec,rinvsq12);
1674
1675             fscal            = felec;
1676
1677             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1678
1679             /* Calculate temporary vectorial force */
1680             tx               = _mm_mul_ps(fscal,dx12);
1681             ty               = _mm_mul_ps(fscal,dy12);
1682             tz               = _mm_mul_ps(fscal,dz12);
1683
1684             /* Update vectorial force */
1685             fix1             = _mm_add_ps(fix1,tx);
1686             fiy1             = _mm_add_ps(fiy1,ty);
1687             fiz1             = _mm_add_ps(fiz1,tz);
1688
1689             fjx2             = _mm_add_ps(fjx2,tx);
1690             fjy2             = _mm_add_ps(fjy2,ty);
1691             fjz2             = _mm_add_ps(fjz2,tz);
1692
1693             /**************************
1694              * CALCULATE INTERACTIONS *
1695              **************************/
1696
1697             /* COULOMB ELECTROSTATICS */
1698             velec            = _mm_mul_ps(qq20,rinv20);
1699             felec            = _mm_mul_ps(velec,rinvsq20);
1700
1701             fscal            = felec;
1702
1703             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1704
1705             /* Calculate temporary vectorial force */
1706             tx               = _mm_mul_ps(fscal,dx20);
1707             ty               = _mm_mul_ps(fscal,dy20);
1708             tz               = _mm_mul_ps(fscal,dz20);
1709
1710             /* Update vectorial force */
1711             fix2             = _mm_add_ps(fix2,tx);
1712             fiy2             = _mm_add_ps(fiy2,ty);
1713             fiz2             = _mm_add_ps(fiz2,tz);
1714
1715             fjx0             = _mm_add_ps(fjx0,tx);
1716             fjy0             = _mm_add_ps(fjy0,ty);
1717             fjz0             = _mm_add_ps(fjz0,tz);
1718
1719             /**************************
1720              * CALCULATE INTERACTIONS *
1721              **************************/
1722
1723             /* COULOMB ELECTROSTATICS */
1724             velec            = _mm_mul_ps(qq21,rinv21);
1725             felec            = _mm_mul_ps(velec,rinvsq21);
1726
1727             fscal            = felec;
1728
1729             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1730
1731             /* Calculate temporary vectorial force */
1732             tx               = _mm_mul_ps(fscal,dx21);
1733             ty               = _mm_mul_ps(fscal,dy21);
1734             tz               = _mm_mul_ps(fscal,dz21);
1735
1736             /* Update vectorial force */
1737             fix2             = _mm_add_ps(fix2,tx);
1738             fiy2             = _mm_add_ps(fiy2,ty);
1739             fiz2             = _mm_add_ps(fiz2,tz);
1740
1741             fjx1             = _mm_add_ps(fjx1,tx);
1742             fjy1             = _mm_add_ps(fjy1,ty);
1743             fjz1             = _mm_add_ps(fjz1,tz);
1744
1745             /**************************
1746              * CALCULATE INTERACTIONS *
1747              **************************/
1748
1749             /* COULOMB ELECTROSTATICS */
1750             velec            = _mm_mul_ps(qq22,rinv22);
1751             felec            = _mm_mul_ps(velec,rinvsq22);
1752
1753             fscal            = felec;
1754
1755             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1756
1757             /* Calculate temporary vectorial force */
1758             tx               = _mm_mul_ps(fscal,dx22);
1759             ty               = _mm_mul_ps(fscal,dy22);
1760             tz               = _mm_mul_ps(fscal,dz22);
1761
1762             /* Update vectorial force */
1763             fix2             = _mm_add_ps(fix2,tx);
1764             fiy2             = _mm_add_ps(fiy2,ty);
1765             fiz2             = _mm_add_ps(fiz2,tz);
1766
1767             fjx2             = _mm_add_ps(fjx2,tx);
1768             fjy2             = _mm_add_ps(fjy2,ty);
1769             fjz2             = _mm_add_ps(fjz2,tz);
1770
1771             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1772             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1773             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1774             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1775
1776             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1777                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1778
1779             /* Inner loop uses 250 flops */
1780         }
1781
1782         /* End of innermost loop */
1783
1784         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1785                                               f+i_coord_offset,fshift+i_shift_offset);
1786
1787         /* Increment number of inner iterations */
1788         inneriter                  += j_index_end - j_index_start;
1789
1790         /* Outer loop uses 18 flops */
1791     }
1792
1793     /* Increment number of outer iterations */
1794     outeriter        += nri;
1795
1796     /* Update outer/inner flops */
1797
1798     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*250);
1799 }