src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_ElecCoul_VdwNone_GeomW3W3_sse4_1_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013, 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 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_sse4_1_single.h"
  50 #include "kernelutil_x86_sse4_1_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW3W3_VF_sse4_1_single
  54  * Electrostatics interaction: Coulomb
  55  * VdW interaction:            None
  56  * Geometry:                   Water3-Water3
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecCoul_VdwNone_GeomW3W3_VF_sse4_1_single
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  83     real             scratch[4*DIM];
  84     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  85     int              vdwioffset0;
  86     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwioffset1;
  88     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  89     int              vdwioffset2;
  90     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  91     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  92     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  93     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  94     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  95     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  96     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  97     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  98     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  99     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 100     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 101     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 102     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 103     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 104     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 105     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 106     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 107     real             *charge;
 108     __m128           dummy_mask,cutoff_mask;
 109     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 110     __m128           one     = _mm_set1_ps(1.0);
 111     __m128           two     = _mm_set1_ps(2.0);
 112     x                = xx[0];
 113     f                = ff[0];
 114
 115     nri              = nlist->nri;
 116     iinr             = nlist->iinr;
 117     jindex           = nlist->jindex;
 118     jjnr             = nlist->jjnr;
 119     shiftidx         = nlist->shift;
 120     gid              = nlist->gid;
 121     shiftvec         = fr->shift_vec[0];
 122     fshift           = fr->fshift[0];
 123     facel            = _mm_set1_ps(fr->epsfac);
 124     charge           = mdatoms->chargeA;
 125
 126     /* Setup water-specific parameters */
 127     inr              = nlist->iinr[0];
 128     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 129     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 130     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 131
 132     jq0              = _mm_set1_ps(charge[inr+0]);
 133     jq1              = _mm_set1_ps(charge[inr+1]);
 134     jq2              = _mm_set1_ps(charge[inr+2]);
 135     qq00             = _mm_mul_ps(iq0,jq0);
 136     qq01             = _mm_mul_ps(iq0,jq1);
 137     qq02             = _mm_mul_ps(iq0,jq2);
 138     qq10             = _mm_mul_ps(iq1,jq0);
 139     qq11             = _mm_mul_ps(iq1,jq1);
 140     qq12             = _mm_mul_ps(iq1,jq2);
 141     qq20             = _mm_mul_ps(iq2,jq0);
 142     qq21             = _mm_mul_ps(iq2,jq1);
 143     qq22             = _mm_mul_ps(iq2,jq2);
 144
 145     /* Avoid stupid compiler warnings */
 146     jnrA = jnrB = jnrC = jnrD = 0;
 147     j_coord_offsetA = 0;
 148     j_coord_offsetB = 0;
 149     j_coord_offsetC = 0;
 150     j_coord_offsetD = 0;
 151
 152     outeriter        = 0;
 153     inneriter        = 0;
 154
 155     for(iidx=0;iidx<4*DIM;iidx++)
 156     {
 157         scratch[iidx] = 0.0;
 158     }
 159
 160     /* Start outer loop over neighborlists */
 161     for(iidx=0; iidx<nri; iidx++)
 162     {
 163         /* Load shift vector for this list */
 164         i_shift_offset   = DIM*shiftidx[iidx];
 165
 166         /* Load limits for loop over neighbors */
 167         j_index_start    = jindex[iidx];
 168         j_index_end      = jindex[iidx+1];
 169
 170         /* Get outer coordinate index */
 171         inr              = iinr[iidx];
 172         i_coord_offset   = DIM*inr;
 173
 174         /* Load i particle coords and add shift vector */
 175         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 176                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 177
 178         fix0             = _mm_setzero_ps();
 179         fiy0             = _mm_setzero_ps();
 180         fiz0             = _mm_setzero_ps();
 181         fix1             = _mm_setzero_ps();
 182         fiy1             = _mm_setzero_ps();
 183         fiz1             = _mm_setzero_ps();
 184         fix2             = _mm_setzero_ps();
 185         fiy2             = _mm_setzero_ps();
 186         fiz2             = _mm_setzero_ps();
 187
 188         /* Reset potential sums */
 189         velecsum         = _mm_setzero_ps();
 190
 191         /* Start inner kernel loop */
 192         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 193         {
 194
 195             /* Get j neighbor index, and coordinate index */
 196             jnrA             = jjnr[jidx];
 197             jnrB             = jjnr[jidx+1];
 198             jnrC             = jjnr[jidx+2];
 199             jnrD             = jjnr[jidx+3];
 200             j_coord_offsetA  = DIM*jnrA;
 201             j_coord_offsetB  = DIM*jnrB;
 202             j_coord_offsetC  = DIM*jnrC;
 203             j_coord_offsetD  = DIM*jnrD;
 204
 205             /* load j atom coordinates */
 206             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 207                                               x+j_coord_offsetC,x+j_coord_offsetD,
 208                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 209
 210             /* Calculate displacement vector */
 211             dx00             = _mm_sub_ps(ix0,jx0);
 212             dy00             = _mm_sub_ps(iy0,jy0);
 213             dz00             = _mm_sub_ps(iz0,jz0);
 214             dx01             = _mm_sub_ps(ix0,jx1);
 215             dy01             = _mm_sub_ps(iy0,jy1);
 216             dz01             = _mm_sub_ps(iz0,jz1);
 217             dx02             = _mm_sub_ps(ix0,jx2);
 218             dy02             = _mm_sub_ps(iy0,jy2);
 219             dz02             = _mm_sub_ps(iz0,jz2);
 220             dx10             = _mm_sub_ps(ix1,jx0);
 221             dy10             = _mm_sub_ps(iy1,jy0);
 222             dz10             = _mm_sub_ps(iz1,jz0);
 223             dx11             = _mm_sub_ps(ix1,jx1);
 224             dy11             = _mm_sub_ps(iy1,jy1);
 225             dz11             = _mm_sub_ps(iz1,jz1);
 226             dx12             = _mm_sub_ps(ix1,jx2);
 227             dy12             = _mm_sub_ps(iy1,jy2);
 228             dz12             = _mm_sub_ps(iz1,jz2);
 229             dx20             = _mm_sub_ps(ix2,jx0);
 230             dy20             = _mm_sub_ps(iy2,jy0);
 231             dz20             = _mm_sub_ps(iz2,jz0);
 232             dx21             = _mm_sub_ps(ix2,jx1);
 233             dy21             = _mm_sub_ps(iy2,jy1);
 234             dz21             = _mm_sub_ps(iz2,jz1);
 235             dx22             = _mm_sub_ps(ix2,jx2);
 236             dy22             = _mm_sub_ps(iy2,jy2);
 237             dz22             = _mm_sub_ps(iz2,jz2);
 238
 239             /* Calculate squared distance and things based on it */
 240             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 241             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 242             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 243             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 244             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 245             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 246             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 247             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 248             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 249
 250             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 251             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 252             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 253             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 254             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 255             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 256             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 257             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 258             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 259
 260             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 261             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 262             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 263             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 264             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 265             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 266             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 267             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 268             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 269
 270             fjx0             = _mm_setzero_ps();
 271             fjy0             = _mm_setzero_ps();
 272             fjz0             = _mm_setzero_ps();
 273             fjx1             = _mm_setzero_ps();
 274             fjy1             = _mm_setzero_ps();
 275             fjz1             = _mm_setzero_ps();
 276             fjx2             = _mm_setzero_ps();
 277             fjy2             = _mm_setzero_ps();
 278             fjz2             = _mm_setzero_ps();
 279
 280             /**************************
 281              * CALCULATE INTERACTIONS *
 282              **************************/
 283
 284             /* COULOMB ELECTROSTATICS */
 285             velec            = _mm_mul_ps(qq00,rinv00);
 286             felec            = _mm_mul_ps(velec,rinvsq00);
 287
 288             /* Update potential sum for this i atom from the interaction with this j atom. */
 289             velecsum         = _mm_add_ps(velecsum,velec);
 290
 291             fscal            = felec;
 292
 293             /* Calculate temporary vectorial force */
 294             tx               = _mm_mul_ps(fscal,dx00);
 295             ty               = _mm_mul_ps(fscal,dy00);
 296             tz               = _mm_mul_ps(fscal,dz00);
 297
 298             /* Update vectorial force */
 299             fix0             = _mm_add_ps(fix0,tx);
 300             fiy0             = _mm_add_ps(fiy0,ty);
 301             fiz0             = _mm_add_ps(fiz0,tz);
 302
 303             fjx0             = _mm_add_ps(fjx0,tx);
 304             fjy0             = _mm_add_ps(fjy0,ty);
 305             fjz0             = _mm_add_ps(fjz0,tz);
 306
 307             /**************************
 308              * CALCULATE INTERACTIONS *
 309              **************************/
 310
 311             /* COULOMB ELECTROSTATICS */
 312             velec            = _mm_mul_ps(qq01,rinv01);
 313             felec            = _mm_mul_ps(velec,rinvsq01);
 314
 315             /* Update potential sum for this i atom from the interaction with this j atom. */
 316             velecsum         = _mm_add_ps(velecsum,velec);
 317
 318             fscal            = felec;
 319
 320             /* Calculate temporary vectorial force */
 321             tx               = _mm_mul_ps(fscal,dx01);
 322             ty               = _mm_mul_ps(fscal,dy01);
 323             tz               = _mm_mul_ps(fscal,dz01);
 324
 325             /* Update vectorial force */
 326             fix0             = _mm_add_ps(fix0,tx);
 327             fiy0             = _mm_add_ps(fiy0,ty);
 328             fiz0             = _mm_add_ps(fiz0,tz);
 329
 330             fjx1             = _mm_add_ps(fjx1,tx);
 331             fjy1             = _mm_add_ps(fjy1,ty);
 332             fjz1             = _mm_add_ps(fjz1,tz);
 333
 334             /**************************
 335              * CALCULATE INTERACTIONS *
 336              **************************/
 337
 338             /* COULOMB ELECTROSTATICS */
 339             velec            = _mm_mul_ps(qq02,rinv02);
 340             felec            = _mm_mul_ps(velec,rinvsq02);
 341
 342             /* Update potential sum for this i atom from the interaction with this j atom. */
 343             velecsum         = _mm_add_ps(velecsum,velec);
 344
 345             fscal            = felec;
 346
 347             /* Calculate temporary vectorial force */
 348             tx               = _mm_mul_ps(fscal,dx02);
 349             ty               = _mm_mul_ps(fscal,dy02);
 350             tz               = _mm_mul_ps(fscal,dz02);
 351
 352             /* Update vectorial force */
 353             fix0             = _mm_add_ps(fix0,tx);
 354             fiy0             = _mm_add_ps(fiy0,ty);
 355             fiz0             = _mm_add_ps(fiz0,tz);
 356
 357             fjx2             = _mm_add_ps(fjx2,tx);
 358             fjy2             = _mm_add_ps(fjy2,ty);
 359             fjz2             = _mm_add_ps(fjz2,tz);
 360
 361             /**************************
 362              * CALCULATE INTERACTIONS *
 363              **************************/
 364
 365             /* COULOMB ELECTROSTATICS */
 366             velec            = _mm_mul_ps(qq10,rinv10);
 367             felec            = _mm_mul_ps(velec,rinvsq10);
 368
 369             /* Update potential sum for this i atom from the interaction with this j atom. */
 370             velecsum         = _mm_add_ps(velecsum,velec);
 371
 372             fscal            = felec;
 373
 374             /* Calculate temporary vectorial force */
 375             tx               = _mm_mul_ps(fscal,dx10);
 376             ty               = _mm_mul_ps(fscal,dy10);
 377             tz               = _mm_mul_ps(fscal,dz10);
 378
 379             /* Update vectorial force */
 380             fix1             = _mm_add_ps(fix1,tx);
 381             fiy1             = _mm_add_ps(fiy1,ty);
 382             fiz1             = _mm_add_ps(fiz1,tz);
 383
 384             fjx0             = _mm_add_ps(fjx0,tx);
 385             fjy0             = _mm_add_ps(fjy0,ty);
 386             fjz0             = _mm_add_ps(fjz0,tz);
 387
 388             /**************************
 389              * CALCULATE INTERACTIONS *
 390              **************************/
 391
 392             /* COULOMB ELECTROSTATICS */
 393             velec            = _mm_mul_ps(qq11,rinv11);
 394             felec            = _mm_mul_ps(velec,rinvsq11);
 395
 396             /* Update potential sum for this i atom from the interaction with this j atom. */
 397             velecsum         = _mm_add_ps(velecsum,velec);
 398
 399             fscal            = felec;
 400
 401             /* Calculate temporary vectorial force */
 402             tx               = _mm_mul_ps(fscal,dx11);
 403             ty               = _mm_mul_ps(fscal,dy11);
 404             tz               = _mm_mul_ps(fscal,dz11);
 405
 406             /* Update vectorial force */
 407             fix1             = _mm_add_ps(fix1,tx);
 408             fiy1             = _mm_add_ps(fiy1,ty);
 409             fiz1             = _mm_add_ps(fiz1,tz);
 410
 411             fjx1             = _mm_add_ps(fjx1,tx);
 412             fjy1             = _mm_add_ps(fjy1,ty);
 413             fjz1             = _mm_add_ps(fjz1,tz);
 414
 415             /**************************
 416              * CALCULATE INTERACTIONS *
 417              **************************/
 418
 419             /* COULOMB ELECTROSTATICS */
 420             velec            = _mm_mul_ps(qq12,rinv12);
 421             felec            = _mm_mul_ps(velec,rinvsq12);
 422
 423             /* Update potential sum for this i atom from the interaction with this j atom. */
 424             velecsum         = _mm_add_ps(velecsum,velec);
 425
 426             fscal            = felec;
 427
 428             /* Calculate temporary vectorial force */
 429             tx               = _mm_mul_ps(fscal,dx12);
 430             ty               = _mm_mul_ps(fscal,dy12);
 431             tz               = _mm_mul_ps(fscal,dz12);
 432
 433             /* Update vectorial force */
 434             fix1             = _mm_add_ps(fix1,tx);
 435             fiy1             = _mm_add_ps(fiy1,ty);
 436             fiz1             = _mm_add_ps(fiz1,tz);
 437
 438             fjx2             = _mm_add_ps(fjx2,tx);
 439             fjy2             = _mm_add_ps(fjy2,ty);
 440             fjz2             = _mm_add_ps(fjz2,tz);
 441
 442             /**************************
 443              * CALCULATE INTERACTIONS *
 444              **************************/
 445
 446             /* COULOMB ELECTROSTATICS */
 447             velec            = _mm_mul_ps(qq20,rinv20);
 448             felec            = _mm_mul_ps(velec,rinvsq20);
 449
 450             /* Update potential sum for this i atom from the interaction with this j atom. */
 451             velecsum         = _mm_add_ps(velecsum,velec);
 452
 453             fscal            = felec;
 454
 455             /* Calculate temporary vectorial force */
 456             tx               = _mm_mul_ps(fscal,dx20);
 457             ty               = _mm_mul_ps(fscal,dy20);
 458             tz               = _mm_mul_ps(fscal,dz20);
 459
 460             /* Update vectorial force */
 461             fix2             = _mm_add_ps(fix2,tx);
 462             fiy2             = _mm_add_ps(fiy2,ty);
 463             fiz2             = _mm_add_ps(fiz2,tz);
 464
 465             fjx0             = _mm_add_ps(fjx0,tx);
 466             fjy0             = _mm_add_ps(fjy0,ty);
 467             fjz0             = _mm_add_ps(fjz0,tz);
 468
 469             /**************************
 470              * CALCULATE INTERACTIONS *
 471              **************************/
 472
 473             /* COULOMB ELECTROSTATICS */
 474             velec            = _mm_mul_ps(qq21,rinv21);
 475             felec            = _mm_mul_ps(velec,rinvsq21);
 476
 477             /* Update potential sum for this i atom from the interaction with this j atom. */
 478             velecsum         = _mm_add_ps(velecsum,velec);
 479
 480             fscal            = felec;
 481
 482             /* Calculate temporary vectorial force */
 483             tx               = _mm_mul_ps(fscal,dx21);
 484             ty               = _mm_mul_ps(fscal,dy21);
 485             tz               = _mm_mul_ps(fscal,dz21);
 486
 487             /* Update vectorial force */
 488             fix2             = _mm_add_ps(fix2,tx);
 489             fiy2             = _mm_add_ps(fiy2,ty);
 490             fiz2             = _mm_add_ps(fiz2,tz);
 491
 492             fjx1             = _mm_add_ps(fjx1,tx);
 493             fjy1             = _mm_add_ps(fjy1,ty);
 494             fjz1             = _mm_add_ps(fjz1,tz);
 495
 496             /**************************
 497              * CALCULATE INTERACTIONS *
 498              **************************/
 499
 500             /* COULOMB ELECTROSTATICS */
 501             velec            = _mm_mul_ps(qq22,rinv22);
 502             felec            = _mm_mul_ps(velec,rinvsq22);
 503
 504             /* Update potential sum for this i atom from the interaction with this j atom. */
 505             velecsum         = _mm_add_ps(velecsum,velec);
 506
 507             fscal            = felec;
 508
 509             /* Calculate temporary vectorial force */
 510             tx               = _mm_mul_ps(fscal,dx22);
 511             ty               = _mm_mul_ps(fscal,dy22);
 512             tz               = _mm_mul_ps(fscal,dz22);
 513
 514             /* Update vectorial force */
 515             fix2             = _mm_add_ps(fix2,tx);
 516             fiy2             = _mm_add_ps(fiy2,ty);
 517             fiz2             = _mm_add_ps(fiz2,tz);
 518
 519             fjx2             = _mm_add_ps(fjx2,tx);
 520             fjy2             = _mm_add_ps(fjy2,ty);
 521             fjz2             = _mm_add_ps(fjz2,tz);
 522
 523             fjptrA             = f+j_coord_offsetA;
 524             fjptrB             = f+j_coord_offsetB;
 525             fjptrC             = f+j_coord_offsetC;
 526             fjptrD             = f+j_coord_offsetD;
 527
 528             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 529                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 530
 531             /* Inner loop uses 252 flops */
 532         }
 533
 534         if(jidx<j_index_end)
 535         {
 536
 537             /* Get j neighbor index, and coordinate index */
 538             jnrlistA         = jjnr[jidx];
 539             jnrlistB         = jjnr[jidx+1];
 540             jnrlistC         = jjnr[jidx+2];
 541             jnrlistD         = jjnr[jidx+3];
 542             /* Sign of each element will be negative for non-real atoms.
 543              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 544              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 545              */
 546             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 547             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 548             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 549             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 550             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 551             j_coord_offsetA  = DIM*jnrA;
 552             j_coord_offsetB  = DIM*jnrB;
 553             j_coord_offsetC  = DIM*jnrC;
 554             j_coord_offsetD  = DIM*jnrD;
 555
 556             /* load j atom coordinates */
 557             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 558                                               x+j_coord_offsetC,x+j_coord_offsetD,
 559                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 560
 561             /* Calculate displacement vector */
 562             dx00             = _mm_sub_ps(ix0,jx0);
 563             dy00             = _mm_sub_ps(iy0,jy0);
 564             dz00             = _mm_sub_ps(iz0,jz0);
 565             dx01             = _mm_sub_ps(ix0,jx1);
 566             dy01             = _mm_sub_ps(iy0,jy1);
 567             dz01             = _mm_sub_ps(iz0,jz1);
 568             dx02             = _mm_sub_ps(ix0,jx2);
 569             dy02             = _mm_sub_ps(iy0,jy2);
 570             dz02             = _mm_sub_ps(iz0,jz2);
 571             dx10             = _mm_sub_ps(ix1,jx0);
 572             dy10             = _mm_sub_ps(iy1,jy0);
 573             dz10             = _mm_sub_ps(iz1,jz0);
 574             dx11             = _mm_sub_ps(ix1,jx1);
 575             dy11             = _mm_sub_ps(iy1,jy1);
 576             dz11             = _mm_sub_ps(iz1,jz1);
 577             dx12             = _mm_sub_ps(ix1,jx2);
 578             dy12             = _mm_sub_ps(iy1,jy2);
 579             dz12             = _mm_sub_ps(iz1,jz2);
 580             dx20             = _mm_sub_ps(ix2,jx0);
 581             dy20             = _mm_sub_ps(iy2,jy0);
 582             dz20             = _mm_sub_ps(iz2,jz0);
 583             dx21             = _mm_sub_ps(ix2,jx1);
 584             dy21             = _mm_sub_ps(iy2,jy1);
 585             dz21             = _mm_sub_ps(iz2,jz1);
 586             dx22             = _mm_sub_ps(ix2,jx2);
 587             dy22             = _mm_sub_ps(iy2,jy2);
 588             dz22             = _mm_sub_ps(iz2,jz2);
 589
 590             /* Calculate squared distance and things based on it */
 591             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 592             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 593             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 594             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 595             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 596             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 597             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 598             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 599             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 600
 601             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 602             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 603             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 604             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 605             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 606             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 607             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 608             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 609             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 610
 611             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 612             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 613             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 614             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 615             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 616             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 617             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 618             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 619             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 620
 621             fjx0             = _mm_setzero_ps();
 622             fjy0             = _mm_setzero_ps();
 623             fjz0             = _mm_setzero_ps();
 624             fjx1             = _mm_setzero_ps();
 625             fjy1             = _mm_setzero_ps();
 626             fjz1             = _mm_setzero_ps();
 627             fjx2             = _mm_setzero_ps();
 628             fjy2             = _mm_setzero_ps();
 629             fjz2             = _mm_setzero_ps();
 630
 631             /**************************
 632              * CALCULATE INTERACTIONS *
 633              **************************/
 634
 635             /* COULOMB ELECTROSTATICS */
 636             velec            = _mm_mul_ps(qq00,rinv00);
 637             felec            = _mm_mul_ps(velec,rinvsq00);
 638
 639             /* Update potential sum for this i atom from the interaction with this j atom. */
 640             velec            = _mm_andnot_ps(dummy_mask,velec);
 641             velecsum         = _mm_add_ps(velecsum,velec);
 642
 643             fscal            = felec;
 644
 645             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 646
 647             /* Calculate temporary vectorial force */
 648             tx               = _mm_mul_ps(fscal,dx00);
 649             ty               = _mm_mul_ps(fscal,dy00);
 650             tz               = _mm_mul_ps(fscal,dz00);
 651
 652             /* Update vectorial force */
 653             fix0             = _mm_add_ps(fix0,tx);
 654             fiy0             = _mm_add_ps(fiy0,ty);
 655             fiz0             = _mm_add_ps(fiz0,tz);
 656
 657             fjx0             = _mm_add_ps(fjx0,tx);
 658             fjy0             = _mm_add_ps(fjy0,ty);
 659             fjz0             = _mm_add_ps(fjz0,tz);
 660
 661             /**************************
 662              * CALCULATE INTERACTIONS *
 663              **************************/
 664
 665             /* COULOMB ELECTROSTATICS */
 666             velec            = _mm_mul_ps(qq01,rinv01);
 667             felec            = _mm_mul_ps(velec,rinvsq01);
 668
 669             /* Update potential sum for this i atom from the interaction with this j atom. */
 670             velec            = _mm_andnot_ps(dummy_mask,velec);
 671             velecsum         = _mm_add_ps(velecsum,velec);
 672
 673             fscal            = felec;
 674
 675             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 676
 677             /* Calculate temporary vectorial force */
 678             tx               = _mm_mul_ps(fscal,dx01);
 679             ty               = _mm_mul_ps(fscal,dy01);
 680             tz               = _mm_mul_ps(fscal,dz01);
 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             fjx1             = _mm_add_ps(fjx1,tx);
 688             fjy1             = _mm_add_ps(fjy1,ty);
 689             fjz1             = _mm_add_ps(fjz1,tz);
 690
 691             /**************************
 692              * CALCULATE INTERACTIONS *
 693              **************************/
 694
 695             /* COULOMB ELECTROSTATICS */
 696             velec            = _mm_mul_ps(qq02,rinv02);
 697             felec            = _mm_mul_ps(velec,rinvsq02);
 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,dx02);
 709             ty               = _mm_mul_ps(fscal,dy02);
 710             tz               = _mm_mul_ps(fscal,dz02);
 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             fjx2             = _mm_add_ps(fjx2,tx);
 718             fjy2             = _mm_add_ps(fjy2,ty);
 719             fjz2             = _mm_add_ps(fjz2,tz);
 720
 721             /**************************
 722              * CALCULATE INTERACTIONS *
 723              **************************/
 724
 725             /* COULOMB ELECTROSTATICS */
 726             velec            = _mm_mul_ps(qq10,rinv10);
 727             felec            = _mm_mul_ps(velec,rinvsq10);
 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,dx10);
 739             ty               = _mm_mul_ps(fscal,dy10);
 740             tz               = _mm_mul_ps(fscal,dz10);
 741
 742             /* Update vectorial force */
 743             fix1             = _mm_add_ps(fix1,tx);
 744             fiy1             = _mm_add_ps(fiy1,ty);
 745             fiz1             = _mm_add_ps(fiz1,tz);
 746
 747             fjx0             = _mm_add_ps(fjx0,tx);
 748             fjy0             = _mm_add_ps(fjy0,ty);
 749             fjz0             = _mm_add_ps(fjz0,tz);
 750
 751             /**************************
 752              * CALCULATE INTERACTIONS *
 753              **************************/
 754
 755             /* COULOMB ELECTROSTATICS */
 756             velec            = _mm_mul_ps(qq11,rinv11);
 757             felec            = _mm_mul_ps(velec,rinvsq11);
 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,dx11);
 769             ty               = _mm_mul_ps(fscal,dy11);
 770             tz               = _mm_mul_ps(fscal,dz11);
 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             fjx1             = _mm_add_ps(fjx1,tx);
 778             fjy1             = _mm_add_ps(fjy1,ty);
 779             fjz1             = _mm_add_ps(fjz1,tz);
 780
 781             /**************************
 782              * CALCULATE INTERACTIONS *
 783              **************************/
 784
 785             /* COULOMB ELECTROSTATICS */
 786             velec            = _mm_mul_ps(qq12,rinv12);
 787             felec            = _mm_mul_ps(velec,rinvsq12);
 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,dx12);
 799             ty               = _mm_mul_ps(fscal,dy12);
 800             tz               = _mm_mul_ps(fscal,dz12);
 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             fjx2             = _mm_add_ps(fjx2,tx);
 808             fjy2             = _mm_add_ps(fjy2,ty);
 809             fjz2             = _mm_add_ps(fjz2,tz);
 810
 811             /**************************
 812              * CALCULATE INTERACTIONS *
 813              **************************/
 814
 815             /* COULOMB ELECTROSTATICS */
 816             velec            = _mm_mul_ps(qq20,rinv20);
 817             felec            = _mm_mul_ps(velec,rinvsq20);
 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,dx20);
 829             ty               = _mm_mul_ps(fscal,dy20);
 830             tz               = _mm_mul_ps(fscal,dz20);
 831
 832             /* Update vectorial force */
 833             fix2             = _mm_add_ps(fix2,tx);
 834             fiy2             = _mm_add_ps(fiy2,ty);
 835             fiz2             = _mm_add_ps(fiz2,tz);
 836
 837             fjx0             = _mm_add_ps(fjx0,tx);
 838             fjy0             = _mm_add_ps(fjy0,ty);
 839             fjz0             = _mm_add_ps(fjz0,tz);
 840
 841             /**************************
 842              * CALCULATE INTERACTIONS *
 843              **************************/
 844
 845             /* COULOMB ELECTROSTATICS */
 846             velec            = _mm_mul_ps(qq21,rinv21);
 847             felec            = _mm_mul_ps(velec,rinvsq21);
 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,dx21);
 859             ty               = _mm_mul_ps(fscal,dy21);
 860             tz               = _mm_mul_ps(fscal,dz21);
 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             fjx1             = _mm_add_ps(fjx1,tx);
 868             fjy1             = _mm_add_ps(fjy1,ty);
 869             fjz1             = _mm_add_ps(fjz1,tz);
 870
 871             /**************************
 872              * CALCULATE INTERACTIONS *
 873              **************************/
 874
 875             /* COULOMB ELECTROSTATICS */
 876             velec            = _mm_mul_ps(qq22,rinv22);
 877             felec            = _mm_mul_ps(velec,rinvsq22);
 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,dx22);
 889             ty               = _mm_mul_ps(fscal,dy22);
 890             tz               = _mm_mul_ps(fscal,dz22);
 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             fjx2             = _mm_add_ps(fjx2,tx);
 898             fjy2             = _mm_add_ps(fjy2,ty);
 899             fjz2             = _mm_add_ps(fjz2,tz);
 900
 901             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 902             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 903             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 904             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 905
 906             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 907                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 908
 909             /* Inner loop uses 252 flops */
 910         }
 911
 912         /* End of innermost loop */
 913
 914         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 915                                               f+i_coord_offset,fshift+i_shift_offset);
 916
 917         ggid                        = gid[iidx];
 918         /* Update potential energies */
 919         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 920
 921         /* Increment number of inner iterations */
 922         inneriter                  += j_index_end - j_index_start;
 923
 924         /* Outer loop uses 19 flops */
 925     }
 926
 927     /* Increment number of outer iterations */
 928     outeriter        += nri;
 929
 930     /* Update outer/inner flops */
 931
 932     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_VF,outeriter*19 + inneriter*252);
 933 }
 934 /*
 935  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW3W3_F_sse4_1_single
 936  * Electrostatics interaction: Coulomb
 937  * VdW interaction:            None
 938  * Geometry:                   Water3-Water3
 939  * Calculate force/pot:        Force
 940  */
 941 void
 942 nb_kernel_ElecCoul_VdwNone_GeomW3W3_F_sse4_1_single
 943                     (t_nblist                    * gmx_restrict       nlist,
 944                      rvec                        * gmx_restrict          xx,
 945                      rvec                        * gmx_restrict          ff,
 946                      t_forcerec                  * gmx_restrict          fr,
 947                      t_mdatoms                   * gmx_restrict     mdatoms,
 948                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 949                      t_nrnb                      * gmx_restrict        nrnb)
 950 {
 951     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 952      * just 0 for non-waters.
 953      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 954      * jnr indices corresponding to data put in the four positions in the SIMD register.
 955      */
 956     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 957     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 958     int              jnrA,jnrB,jnrC,jnrD;
 959     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 960     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 961     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 962     real             rcutoff_scalar;
 963     real             *shiftvec,*fshift,*x,*f;
 964     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 965     real             scratch[4*DIM];
 966     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 967     int              vdwioffset0;
 968     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 969     int              vdwioffset1;
 970     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 971     int              vdwioffset2;
 972     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 973     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 974     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 975     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
 976     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 977     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
 978     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 979     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 980     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 981     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 982     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 983     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 984     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 985     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 986     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 987     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 988     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 989     real             *charge;
 990     __m128           dummy_mask,cutoff_mask;
 991     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 992     __m128           one     = _mm_set1_ps(1.0);
 993     __m128           two     = _mm_set1_ps(2.0);
 994     x                = xx[0];
 995     f                = ff[0];
 996
 997     nri              = nlist->nri;
 998     iinr             = nlist->iinr;
 999     jindex           = nlist->jindex;
1000     jjnr             = nlist->jjnr;
1001     shiftidx         = nlist->shift;
1002     gid              = nlist->gid;
1003     shiftvec         = fr->shift_vec[0];
1004     fshift           = fr->fshift[0];
1005     facel            = _mm_set1_ps(fr->epsfac);
1006     charge           = mdatoms->chargeA;
1007
1008     /* Setup water-specific parameters */
1009     inr              = nlist->iinr[0];
1010     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1011     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1012     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1013
1014     jq0              = _mm_set1_ps(charge[inr+0]);
1015     jq1              = _mm_set1_ps(charge[inr+1]);
1016     jq2              = _mm_set1_ps(charge[inr+2]);
1017     qq00             = _mm_mul_ps(iq0,jq0);
1018     qq01             = _mm_mul_ps(iq0,jq1);
1019     qq02             = _mm_mul_ps(iq0,jq2);
1020     qq10             = _mm_mul_ps(iq1,jq0);
1021     qq11             = _mm_mul_ps(iq1,jq1);
1022     qq12             = _mm_mul_ps(iq1,jq2);
1023     qq20             = _mm_mul_ps(iq2,jq0);
1024     qq21             = _mm_mul_ps(iq2,jq1);
1025     qq22             = _mm_mul_ps(iq2,jq2);
1026
1027     /* Avoid stupid compiler warnings */
1028     jnrA = jnrB = jnrC = jnrD = 0;
1029     j_coord_offsetA = 0;
1030     j_coord_offsetB = 0;
1031     j_coord_offsetC = 0;
1032     j_coord_offsetD = 0;
1033
1034     outeriter        = 0;
1035     inneriter        = 0;
1036
1037     for(iidx=0;iidx<4*DIM;iidx++)
1038     {
1039         scratch[iidx] = 0.0;
1040     }
1041
1042     /* Start outer loop over neighborlists */
1043     for(iidx=0; iidx<nri; iidx++)
1044     {
1045         /* Load shift vector for this list */
1046         i_shift_offset   = DIM*shiftidx[iidx];
1047
1048         /* Load limits for loop over neighbors */
1049         j_index_start    = jindex[iidx];
1050         j_index_end      = jindex[iidx+1];
1051
1052         /* Get outer coordinate index */
1053         inr              = iinr[iidx];
1054         i_coord_offset   = DIM*inr;
1055
1056         /* Load i particle coords and add shift vector */
1057         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1058                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1059
1060         fix0             = _mm_setzero_ps();
1061         fiy0             = _mm_setzero_ps();
1062         fiz0             = _mm_setzero_ps();
1063         fix1             = _mm_setzero_ps();
1064         fiy1             = _mm_setzero_ps();
1065         fiz1             = _mm_setzero_ps();
1066         fix2             = _mm_setzero_ps();
1067         fiy2             = _mm_setzero_ps();
1068         fiz2             = _mm_setzero_ps();
1069
1070         /* Start inner kernel loop */
1071         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1072         {
1073
1074             /* Get j neighbor index, and coordinate index */
1075             jnrA             = jjnr[jidx];
1076             jnrB             = jjnr[jidx+1];
1077             jnrC             = jjnr[jidx+2];
1078             jnrD             = jjnr[jidx+3];
1079             j_coord_offsetA  = DIM*jnrA;
1080             j_coord_offsetB  = DIM*jnrB;
1081             j_coord_offsetC  = DIM*jnrC;
1082             j_coord_offsetD  = DIM*jnrD;
1083
1084             /* load j atom coordinates */
1085             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1086                                               x+j_coord_offsetC,x+j_coord_offsetD,
1087                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1088
1089             /* Calculate displacement vector */
1090             dx00             = _mm_sub_ps(ix0,jx0);
1091             dy00             = _mm_sub_ps(iy0,jy0);
1092             dz00             = _mm_sub_ps(iz0,jz0);
1093             dx01             = _mm_sub_ps(ix0,jx1);
1094             dy01             = _mm_sub_ps(iy0,jy1);
1095             dz01             = _mm_sub_ps(iz0,jz1);
1096             dx02             = _mm_sub_ps(ix0,jx2);
1097             dy02             = _mm_sub_ps(iy0,jy2);
1098             dz02             = _mm_sub_ps(iz0,jz2);
1099             dx10             = _mm_sub_ps(ix1,jx0);
1100             dy10             = _mm_sub_ps(iy1,jy0);
1101             dz10             = _mm_sub_ps(iz1,jz0);
1102             dx11             = _mm_sub_ps(ix1,jx1);
1103             dy11             = _mm_sub_ps(iy1,jy1);
1104             dz11             = _mm_sub_ps(iz1,jz1);
1105             dx12             = _mm_sub_ps(ix1,jx2);
1106             dy12             = _mm_sub_ps(iy1,jy2);
1107             dz12             = _mm_sub_ps(iz1,jz2);
1108             dx20             = _mm_sub_ps(ix2,jx0);
1109             dy20             = _mm_sub_ps(iy2,jy0);
1110             dz20             = _mm_sub_ps(iz2,jz0);
1111             dx21             = _mm_sub_ps(ix2,jx1);
1112             dy21             = _mm_sub_ps(iy2,jy1);
1113             dz21             = _mm_sub_ps(iz2,jz1);
1114             dx22             = _mm_sub_ps(ix2,jx2);
1115             dy22             = _mm_sub_ps(iy2,jy2);
1116             dz22             = _mm_sub_ps(iz2,jz2);
1117
1118             /* Calculate squared distance and things based on it */
1119             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1120             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1121             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1122             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1123             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1124             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1125             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1126             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1127             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1128
1129             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1130             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1131             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1132             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1133             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1134             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1135             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1136             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1137             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1138
1139             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1140             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1141             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1142             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1143             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1144             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1145             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1146             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1147             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1148
1149             fjx0             = _mm_setzero_ps();
1150             fjy0             = _mm_setzero_ps();
1151             fjz0             = _mm_setzero_ps();
1152             fjx1             = _mm_setzero_ps();
1153             fjy1             = _mm_setzero_ps();
1154             fjz1             = _mm_setzero_ps();
1155             fjx2             = _mm_setzero_ps();
1156             fjy2             = _mm_setzero_ps();
1157             fjz2             = _mm_setzero_ps();
1158
1159             /**************************
1160              * CALCULATE INTERACTIONS *
1161              **************************/
1162
1163             /* COULOMB ELECTROSTATICS */
1164             velec            = _mm_mul_ps(qq00,rinv00);
1165             felec            = _mm_mul_ps(velec,rinvsq00);
1166
1167             fscal            = felec;
1168
1169             /* Calculate temporary vectorial force */
1170             tx               = _mm_mul_ps(fscal,dx00);
1171             ty               = _mm_mul_ps(fscal,dy00);
1172             tz               = _mm_mul_ps(fscal,dz00);
1173
1174             /* Update vectorial force */
1175             fix0             = _mm_add_ps(fix0,tx);
1176             fiy0             = _mm_add_ps(fiy0,ty);
1177             fiz0             = _mm_add_ps(fiz0,tz);
1178
1179             fjx0             = _mm_add_ps(fjx0,tx);
1180             fjy0             = _mm_add_ps(fjy0,ty);
1181             fjz0             = _mm_add_ps(fjz0,tz);
1182
1183             /**************************
1184              * CALCULATE INTERACTIONS *
1185              **************************/
1186
1187             /* COULOMB ELECTROSTATICS */
1188             velec            = _mm_mul_ps(qq01,rinv01);
1189             felec            = _mm_mul_ps(velec,rinvsq01);
1190
1191             fscal            = felec;
1192
1193             /* Calculate temporary vectorial force */
1194             tx               = _mm_mul_ps(fscal,dx01);
1195             ty               = _mm_mul_ps(fscal,dy01);
1196             tz               = _mm_mul_ps(fscal,dz01);
1197
1198             /* Update vectorial force */
1199             fix0             = _mm_add_ps(fix0,tx);
1200             fiy0             = _mm_add_ps(fiy0,ty);
1201             fiz0             = _mm_add_ps(fiz0,tz);
1202
1203             fjx1             = _mm_add_ps(fjx1,tx);
1204             fjy1             = _mm_add_ps(fjy1,ty);
1205             fjz1             = _mm_add_ps(fjz1,tz);
1206
1207             /**************************
1208              * CALCULATE INTERACTIONS *
1209              **************************/
1210
1211             /* COULOMB ELECTROSTATICS */
1212             velec            = _mm_mul_ps(qq02,rinv02);
1213             felec            = _mm_mul_ps(velec,rinvsq02);
1214
1215             fscal            = felec;
1216
1217             /* Calculate temporary vectorial force */
1218             tx               = _mm_mul_ps(fscal,dx02);
1219             ty               = _mm_mul_ps(fscal,dy02);
1220             tz               = _mm_mul_ps(fscal,dz02);
1221
1222             /* Update vectorial force */
1223             fix0             = _mm_add_ps(fix0,tx);
1224             fiy0             = _mm_add_ps(fiy0,ty);
1225             fiz0             = _mm_add_ps(fiz0,tz);
1226
1227             fjx2             = _mm_add_ps(fjx2,tx);
1228             fjy2             = _mm_add_ps(fjy2,ty);
1229             fjz2             = _mm_add_ps(fjz2,tz);
1230
1231             /**************************
1232              * CALCULATE INTERACTIONS *
1233              **************************/
1234
1235             /* COULOMB ELECTROSTATICS */
1236             velec            = _mm_mul_ps(qq10,rinv10);
1237             felec            = _mm_mul_ps(velec,rinvsq10);
1238
1239             fscal            = felec;
1240
1241             /* Calculate temporary vectorial force */
1242             tx               = _mm_mul_ps(fscal,dx10);
1243             ty               = _mm_mul_ps(fscal,dy10);
1244             tz               = _mm_mul_ps(fscal,dz10);
1245
1246             /* Update vectorial force */
1247             fix1             = _mm_add_ps(fix1,tx);
1248             fiy1             = _mm_add_ps(fiy1,ty);
1249             fiz1             = _mm_add_ps(fiz1,tz);
1250
1251             fjx0             = _mm_add_ps(fjx0,tx);
1252             fjy0             = _mm_add_ps(fjy0,ty);
1253             fjz0             = _mm_add_ps(fjz0,tz);
1254
1255             /**************************
1256              * CALCULATE INTERACTIONS *
1257              **************************/
1258
1259             /* COULOMB ELECTROSTATICS */
1260             velec            = _mm_mul_ps(qq11,rinv11);
1261             felec            = _mm_mul_ps(velec,rinvsq11);
1262
1263             fscal            = felec;
1264
1265             /* Calculate temporary vectorial force */
1266             tx               = _mm_mul_ps(fscal,dx11);
1267             ty               = _mm_mul_ps(fscal,dy11);
1268             tz               = _mm_mul_ps(fscal,dz11);
1269
1270             /* Update vectorial force */
1271             fix1             = _mm_add_ps(fix1,tx);
1272             fiy1             = _mm_add_ps(fiy1,ty);
1273             fiz1             = _mm_add_ps(fiz1,tz);
1274
1275             fjx1             = _mm_add_ps(fjx1,tx);
1276             fjy1             = _mm_add_ps(fjy1,ty);
1277             fjz1             = _mm_add_ps(fjz1,tz);
1278
1279             /**************************
1280              * CALCULATE INTERACTIONS *
1281              **************************/
1282
1283             /* COULOMB ELECTROSTATICS */
1284             velec            = _mm_mul_ps(qq12,rinv12);
1285             felec            = _mm_mul_ps(velec,rinvsq12);
1286
1287             fscal            = felec;
1288
1289             /* Calculate temporary vectorial force */
1290             tx               = _mm_mul_ps(fscal,dx12);
1291             ty               = _mm_mul_ps(fscal,dy12);
1292             tz               = _mm_mul_ps(fscal,dz12);
1293
1294             /* Update vectorial force */
1295             fix1             = _mm_add_ps(fix1,tx);
1296             fiy1             = _mm_add_ps(fiy1,ty);
1297             fiz1             = _mm_add_ps(fiz1,tz);
1298
1299             fjx2             = _mm_add_ps(fjx2,tx);
1300             fjy2             = _mm_add_ps(fjy2,ty);
1301             fjz2             = _mm_add_ps(fjz2,tz);
1302
1303             /**************************
1304              * CALCULATE INTERACTIONS *
1305              **************************/
1306
1307             /* COULOMB ELECTROSTATICS */
1308             velec            = _mm_mul_ps(qq20,rinv20);
1309             felec            = _mm_mul_ps(velec,rinvsq20);
1310
1311             fscal            = felec;
1312
1313             /* Calculate temporary vectorial force */
1314             tx               = _mm_mul_ps(fscal,dx20);
1315             ty               = _mm_mul_ps(fscal,dy20);
1316             tz               = _mm_mul_ps(fscal,dz20);
1317
1318             /* Update vectorial force */
1319             fix2             = _mm_add_ps(fix2,tx);
1320             fiy2             = _mm_add_ps(fiy2,ty);
1321             fiz2             = _mm_add_ps(fiz2,tz);
1322
1323             fjx0             = _mm_add_ps(fjx0,tx);
1324             fjy0             = _mm_add_ps(fjy0,ty);
1325             fjz0             = _mm_add_ps(fjz0,tz);
1326
1327             /**************************
1328              * CALCULATE INTERACTIONS *
1329              **************************/
1330
1331             /* COULOMB ELECTROSTATICS */
1332             velec            = _mm_mul_ps(qq21,rinv21);
1333             felec            = _mm_mul_ps(velec,rinvsq21);
1334
1335             fscal            = felec;
1336
1337             /* Calculate temporary vectorial force */
1338             tx               = _mm_mul_ps(fscal,dx21);
1339             ty               = _mm_mul_ps(fscal,dy21);
1340             tz               = _mm_mul_ps(fscal,dz21);
1341
1342             /* Update vectorial force */
1343             fix2             = _mm_add_ps(fix2,tx);
1344             fiy2             = _mm_add_ps(fiy2,ty);
1345             fiz2             = _mm_add_ps(fiz2,tz);
1346
1347             fjx1             = _mm_add_ps(fjx1,tx);
1348             fjy1             = _mm_add_ps(fjy1,ty);
1349             fjz1             = _mm_add_ps(fjz1,tz);
1350
1351             /**************************
1352              * CALCULATE INTERACTIONS *
1353              **************************/
1354
1355             /* COULOMB ELECTROSTATICS */
1356             velec            = _mm_mul_ps(qq22,rinv22);
1357             felec            = _mm_mul_ps(velec,rinvsq22);
1358
1359             fscal            = felec;
1360
1361             /* Calculate temporary vectorial force */
1362             tx               = _mm_mul_ps(fscal,dx22);
1363             ty               = _mm_mul_ps(fscal,dy22);
1364             tz               = _mm_mul_ps(fscal,dz22);
1365
1366             /* Update vectorial force */
1367             fix2             = _mm_add_ps(fix2,tx);
1368             fiy2             = _mm_add_ps(fiy2,ty);
1369             fiz2             = _mm_add_ps(fiz2,tz);
1370
1371             fjx2             = _mm_add_ps(fjx2,tx);
1372             fjy2             = _mm_add_ps(fjy2,ty);
1373             fjz2             = _mm_add_ps(fjz2,tz);
1374
1375             fjptrA             = f+j_coord_offsetA;
1376             fjptrB             = f+j_coord_offsetB;
1377             fjptrC             = f+j_coord_offsetC;
1378             fjptrD             = f+j_coord_offsetD;
1379
1380             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1381                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1382
1383             /* Inner loop uses 243 flops */
1384         }
1385
1386         if(jidx<j_index_end)
1387         {
1388
1389             /* Get j neighbor index, and coordinate index */
1390             jnrlistA         = jjnr[jidx];
1391             jnrlistB         = jjnr[jidx+1];
1392             jnrlistC         = jjnr[jidx+2];
1393             jnrlistD         = jjnr[jidx+3];
1394             /* Sign of each element will be negative for non-real atoms.
1395              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1396              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1397              */
1398             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1399             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1400             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1401             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1402             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1403             j_coord_offsetA  = DIM*jnrA;
1404             j_coord_offsetB  = DIM*jnrB;
1405             j_coord_offsetC  = DIM*jnrC;
1406             j_coord_offsetD  = DIM*jnrD;
1407
1408             /* load j atom coordinates */
1409             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1410                                               x+j_coord_offsetC,x+j_coord_offsetD,
1411                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1412
1413             /* Calculate displacement vector */
1414             dx00             = _mm_sub_ps(ix0,jx0);
1415             dy00             = _mm_sub_ps(iy0,jy0);
1416             dz00             = _mm_sub_ps(iz0,jz0);
1417             dx01             = _mm_sub_ps(ix0,jx1);
1418             dy01             = _mm_sub_ps(iy0,jy1);
1419             dz01             = _mm_sub_ps(iz0,jz1);
1420             dx02             = _mm_sub_ps(ix0,jx2);
1421             dy02             = _mm_sub_ps(iy0,jy2);
1422             dz02             = _mm_sub_ps(iz0,jz2);
1423             dx10             = _mm_sub_ps(ix1,jx0);
1424             dy10             = _mm_sub_ps(iy1,jy0);
1425             dz10             = _mm_sub_ps(iz1,jz0);
1426             dx11             = _mm_sub_ps(ix1,jx1);
1427             dy11             = _mm_sub_ps(iy1,jy1);
1428             dz11             = _mm_sub_ps(iz1,jz1);
1429             dx12             = _mm_sub_ps(ix1,jx2);
1430             dy12             = _mm_sub_ps(iy1,jy2);
1431             dz12             = _mm_sub_ps(iz1,jz2);
1432             dx20             = _mm_sub_ps(ix2,jx0);
1433             dy20             = _mm_sub_ps(iy2,jy0);
1434             dz20             = _mm_sub_ps(iz2,jz0);
1435             dx21             = _mm_sub_ps(ix2,jx1);
1436             dy21             = _mm_sub_ps(iy2,jy1);
1437             dz21             = _mm_sub_ps(iz2,jz1);
1438             dx22             = _mm_sub_ps(ix2,jx2);
1439             dy22             = _mm_sub_ps(iy2,jy2);
1440             dz22             = _mm_sub_ps(iz2,jz2);
1441
1442             /* Calculate squared distance and things based on it */
1443             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1444             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1445             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1446             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1447             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1448             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1449             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1450             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1451             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1452
1453             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1454             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1455             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1456             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1457             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1458             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1459             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1460             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1461             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1462
1463             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1464             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1465             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1466             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1467             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1468             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1469             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1470             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1471             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1472
1473             fjx0             = _mm_setzero_ps();
1474             fjy0             = _mm_setzero_ps();
1475             fjz0             = _mm_setzero_ps();
1476             fjx1             = _mm_setzero_ps();
1477             fjy1             = _mm_setzero_ps();
1478             fjz1             = _mm_setzero_ps();
1479             fjx2             = _mm_setzero_ps();
1480             fjy2             = _mm_setzero_ps();
1481             fjz2             = _mm_setzero_ps();
1482
1483             /**************************
1484              * CALCULATE INTERACTIONS *
1485              **************************/
1486
1487             /* COULOMB ELECTROSTATICS */
1488             velec            = _mm_mul_ps(qq00,rinv00);
1489             felec            = _mm_mul_ps(velec,rinvsq00);
1490
1491             fscal            = felec;
1492
1493             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1494
1495             /* Calculate temporary vectorial force */
1496             tx               = _mm_mul_ps(fscal,dx00);
1497             ty               = _mm_mul_ps(fscal,dy00);
1498             tz               = _mm_mul_ps(fscal,dz00);
1499
1500             /* Update vectorial force */
1501             fix0             = _mm_add_ps(fix0,tx);
1502             fiy0             = _mm_add_ps(fiy0,ty);
1503             fiz0             = _mm_add_ps(fiz0,tz);
1504
1505             fjx0             = _mm_add_ps(fjx0,tx);
1506             fjy0             = _mm_add_ps(fjy0,ty);
1507             fjz0             = _mm_add_ps(fjz0,tz);
1508
1509             /**************************
1510              * CALCULATE INTERACTIONS *
1511              **************************/
1512
1513             /* COULOMB ELECTROSTATICS */
1514             velec            = _mm_mul_ps(qq01,rinv01);
1515             felec            = _mm_mul_ps(velec,rinvsq01);
1516
1517             fscal            = felec;
1518
1519             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1520
1521             /* Calculate temporary vectorial force */
1522             tx               = _mm_mul_ps(fscal,dx01);
1523             ty               = _mm_mul_ps(fscal,dy01);
1524             tz               = _mm_mul_ps(fscal,dz01);
1525
1526             /* Update vectorial force */
1527             fix0             = _mm_add_ps(fix0,tx);
1528             fiy0             = _mm_add_ps(fiy0,ty);
1529             fiz0             = _mm_add_ps(fiz0,tz);
1530
1531             fjx1             = _mm_add_ps(fjx1,tx);
1532             fjy1             = _mm_add_ps(fjy1,ty);
1533             fjz1             = _mm_add_ps(fjz1,tz);
1534
1535             /**************************
1536              * CALCULATE INTERACTIONS *
1537              **************************/
1538
1539             /* COULOMB ELECTROSTATICS */
1540             velec            = _mm_mul_ps(qq02,rinv02);
1541             felec            = _mm_mul_ps(velec,rinvsq02);
1542
1543             fscal            = felec;
1544
1545             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1546
1547             /* Calculate temporary vectorial force */
1548             tx               = _mm_mul_ps(fscal,dx02);
1549             ty               = _mm_mul_ps(fscal,dy02);
1550             tz               = _mm_mul_ps(fscal,dz02);
1551
1552             /* Update vectorial force */
1553             fix0             = _mm_add_ps(fix0,tx);
1554             fiy0             = _mm_add_ps(fiy0,ty);
1555             fiz0             = _mm_add_ps(fiz0,tz);
1556
1557             fjx2             = _mm_add_ps(fjx2,tx);
1558             fjy2             = _mm_add_ps(fjy2,ty);
1559             fjz2             = _mm_add_ps(fjz2,tz);
1560
1561             /**************************
1562              * CALCULATE INTERACTIONS *
1563              **************************/
1564
1565             /* COULOMB ELECTROSTATICS */
1566             velec            = _mm_mul_ps(qq10,rinv10);
1567             felec            = _mm_mul_ps(velec,rinvsq10);
1568
1569             fscal            = felec;
1570
1571             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1572
1573             /* Calculate temporary vectorial force */
1574             tx               = _mm_mul_ps(fscal,dx10);
1575             ty               = _mm_mul_ps(fscal,dy10);
1576             tz               = _mm_mul_ps(fscal,dz10);
1577
1578             /* Update vectorial force */
1579             fix1             = _mm_add_ps(fix1,tx);
1580             fiy1             = _mm_add_ps(fiy1,ty);
1581             fiz1             = _mm_add_ps(fiz1,tz);
1582
1583             fjx0             = _mm_add_ps(fjx0,tx);
1584             fjy0             = _mm_add_ps(fjy0,ty);
1585             fjz0             = _mm_add_ps(fjz0,tz);
1586
1587             /**************************
1588              * CALCULATE INTERACTIONS *
1589              **************************/
1590
1591             /* COULOMB ELECTROSTATICS */
1592             velec            = _mm_mul_ps(qq11,rinv11);
1593             felec            = _mm_mul_ps(velec,rinvsq11);
1594
1595             fscal            = felec;
1596
1597             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1598
1599             /* Calculate temporary vectorial force */
1600             tx               = _mm_mul_ps(fscal,dx11);
1601             ty               = _mm_mul_ps(fscal,dy11);
1602             tz               = _mm_mul_ps(fscal,dz11);
1603
1604             /* Update vectorial force */
1605             fix1             = _mm_add_ps(fix1,tx);
1606             fiy1             = _mm_add_ps(fiy1,ty);
1607             fiz1             = _mm_add_ps(fiz1,tz);
1608
1609             fjx1             = _mm_add_ps(fjx1,tx);
1610             fjy1             = _mm_add_ps(fjy1,ty);
1611             fjz1             = _mm_add_ps(fjz1,tz);
1612
1613             /**************************
1614              * CALCULATE INTERACTIONS *
1615              **************************/
1616
1617             /* COULOMB ELECTROSTATICS */
1618             velec            = _mm_mul_ps(qq12,rinv12);
1619             felec            = _mm_mul_ps(velec,rinvsq12);
1620
1621             fscal            = felec;
1622
1623             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1624
1625             /* Calculate temporary vectorial force */
1626             tx               = _mm_mul_ps(fscal,dx12);
1627             ty               = _mm_mul_ps(fscal,dy12);
1628             tz               = _mm_mul_ps(fscal,dz12);
1629
1630             /* Update vectorial force */
1631             fix1             = _mm_add_ps(fix1,tx);
1632             fiy1             = _mm_add_ps(fiy1,ty);
1633             fiz1             = _mm_add_ps(fiz1,tz);
1634
1635             fjx2             = _mm_add_ps(fjx2,tx);
1636             fjy2             = _mm_add_ps(fjy2,ty);
1637             fjz2             = _mm_add_ps(fjz2,tz);
1638
1639             /**************************
1640              * CALCULATE INTERACTIONS *
1641              **************************/
1642
1643             /* COULOMB ELECTROSTATICS */
1644             velec            = _mm_mul_ps(qq20,rinv20);
1645             felec            = _mm_mul_ps(velec,rinvsq20);
1646
1647             fscal            = felec;
1648
1649             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1650
1651             /* Calculate temporary vectorial force */
1652             tx               = _mm_mul_ps(fscal,dx20);
1653             ty               = _mm_mul_ps(fscal,dy20);
1654             tz               = _mm_mul_ps(fscal,dz20);
1655
1656             /* Update vectorial force */
1657             fix2             = _mm_add_ps(fix2,tx);
1658             fiy2             = _mm_add_ps(fiy2,ty);
1659             fiz2             = _mm_add_ps(fiz2,tz);
1660
1661             fjx0             = _mm_add_ps(fjx0,tx);
1662             fjy0             = _mm_add_ps(fjy0,ty);
1663             fjz0             = _mm_add_ps(fjz0,tz);
1664
1665             /**************************
1666              * CALCULATE INTERACTIONS *
1667              **************************/
1668
1669             /* COULOMB ELECTROSTATICS */
1670             velec            = _mm_mul_ps(qq21,rinv21);
1671             felec            = _mm_mul_ps(velec,rinvsq21);
1672
1673             fscal            = felec;
1674
1675             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1676
1677             /* Calculate temporary vectorial force */
1678             tx               = _mm_mul_ps(fscal,dx21);
1679             ty               = _mm_mul_ps(fscal,dy21);
1680             tz               = _mm_mul_ps(fscal,dz21);
1681
1682             /* Update vectorial force */
1683             fix2             = _mm_add_ps(fix2,tx);
1684             fiy2             = _mm_add_ps(fiy2,ty);
1685             fiz2             = _mm_add_ps(fiz2,tz);
1686
1687             fjx1             = _mm_add_ps(fjx1,tx);
1688             fjy1             = _mm_add_ps(fjy1,ty);
1689             fjz1             = _mm_add_ps(fjz1,tz);
1690
1691             /**************************
1692              * CALCULATE INTERACTIONS *
1693              **************************/
1694
1695             /* COULOMB ELECTROSTATICS */
1696             velec            = _mm_mul_ps(qq22,rinv22);
1697             felec            = _mm_mul_ps(velec,rinvsq22);
1698
1699             fscal            = felec;
1700
1701             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1702
1703             /* Calculate temporary vectorial force */
1704             tx               = _mm_mul_ps(fscal,dx22);
1705             ty               = _mm_mul_ps(fscal,dy22);
1706             tz               = _mm_mul_ps(fscal,dz22);
1707
1708             /* Update vectorial force */
1709             fix2             = _mm_add_ps(fix2,tx);
1710             fiy2             = _mm_add_ps(fiy2,ty);
1711             fiz2             = _mm_add_ps(fiz2,tz);
1712
1713             fjx2             = _mm_add_ps(fjx2,tx);
1714             fjy2             = _mm_add_ps(fjy2,ty);
1715             fjz2             = _mm_add_ps(fjz2,tz);
1716
1717             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1718             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1719             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1720             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1721
1722             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1723                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1724
1725             /* Inner loop uses 243 flops */
1726         }
1727
1728         /* End of innermost loop */
1729
1730         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1731                                               f+i_coord_offset,fshift+i_shift_offset);
1732
1733         /* Increment number of inner iterations */
1734         inneriter                  += j_index_end - j_index_start;
1735
1736         /* Outer loop uses 18 flops */
1737     }
1738
1739     /* Increment number of outer iterations */
1740     outeriter        += nri;
1741
1742     /* Update outer/inner flops */
1743
1744     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_F,outeriter*18 + inneriter*243);
1745 }