src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecCoul_VdwNone_GeomW4W4_sse2_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS sse2_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_sse2_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW4W4_VF_sse2_single
  51  * Electrostatics interaction: Coulomb
  52  * VdW interaction:            None
  53  * Geometry:                   Water4-Water4
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecCoul_VdwNone_GeomW4W4_VF_sse2_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              vdwioffset1;
  83     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  84     int              vdwioffset2;
  85     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  86     int              vdwioffset3;
  87     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  88     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  89     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  90     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  91     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  92     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  93     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  94     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  95     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  96     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  97     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  98     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  99     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 100     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 101     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 102     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 103     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 104     real             *charge;
 105     __m128           dummy_mask,cutoff_mask;
 106     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 107     __m128           one     = _mm_set1_ps(1.0);
 108     __m128           two     = _mm_set1_ps(2.0);
 109     x                = xx[0];
 110     f                = ff[0];
 111
 112     nri              = nlist->nri;
 113     iinr             = nlist->iinr;
 114     jindex           = nlist->jindex;
 115     jjnr             = nlist->jjnr;
 116     shiftidx         = nlist->shift;
 117     gid              = nlist->gid;
 118     shiftvec         = fr->shift_vec[0];
 119     fshift           = fr->fshift[0];
 120     facel            = _mm_set1_ps(fr->ic->epsfac);
 121     charge           = mdatoms->chargeA;
 122
 123     /* Setup water-specific parameters */
 124     inr              = nlist->iinr[0];
 125     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 126     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 127     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 128
 129     jq1              = _mm_set1_ps(charge[inr+1]);
 130     jq2              = _mm_set1_ps(charge[inr+2]);
 131     jq3              = _mm_set1_ps(charge[inr+3]);
 132     qq11             = _mm_mul_ps(iq1,jq1);
 133     qq12             = _mm_mul_ps(iq1,jq2);
 134     qq13             = _mm_mul_ps(iq1,jq3);
 135     qq21             = _mm_mul_ps(iq2,jq1);
 136     qq22             = _mm_mul_ps(iq2,jq2);
 137     qq23             = _mm_mul_ps(iq2,jq3);
 138     qq31             = _mm_mul_ps(iq3,jq1);
 139     qq32             = _mm_mul_ps(iq3,jq2);
 140     qq33             = _mm_mul_ps(iq3,jq3);
 141
 142     /* Avoid stupid compiler warnings */
 143     jnrA = jnrB = jnrC = jnrD = 0;
 144     j_coord_offsetA = 0;
 145     j_coord_offsetB = 0;
 146     j_coord_offsetC = 0;
 147     j_coord_offsetD = 0;
 148
 149     outeriter        = 0;
 150     inneriter        = 0;
 151
 152     for(iidx=0;iidx<4*DIM;iidx++)
 153     {
 154         scratch[iidx] = 0.0;
 155     }
 156
 157     /* Start outer loop over neighborlists */
 158     for(iidx=0; iidx<nri; iidx++)
 159     {
 160         /* Load shift vector for this list */
 161         i_shift_offset   = DIM*shiftidx[iidx];
 162
 163         /* Load limits for loop over neighbors */
 164         j_index_start    = jindex[iidx];
 165         j_index_end      = jindex[iidx+1];
 166
 167         /* Get outer coordinate index */
 168         inr              = iinr[iidx];
 169         i_coord_offset   = DIM*inr;
 170
 171         /* Load i particle coords and add shift vector */
 172         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 173                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 174
 175         fix1             = _mm_setzero_ps();
 176         fiy1             = _mm_setzero_ps();
 177         fiz1             = _mm_setzero_ps();
 178         fix2             = _mm_setzero_ps();
 179         fiy2             = _mm_setzero_ps();
 180         fiz2             = _mm_setzero_ps();
 181         fix3             = _mm_setzero_ps();
 182         fiy3             = _mm_setzero_ps();
 183         fiz3             = _mm_setzero_ps();
 184
 185         /* Reset potential sums */
 186         velecsum         = _mm_setzero_ps();
 187
 188         /* Start inner kernel loop */
 189         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 190         {
 191
 192             /* Get j neighbor index, and coordinate index */
 193             jnrA             = jjnr[jidx];
 194             jnrB             = jjnr[jidx+1];
 195             jnrC             = jjnr[jidx+2];
 196             jnrD             = jjnr[jidx+3];
 197             j_coord_offsetA  = DIM*jnrA;
 198             j_coord_offsetB  = DIM*jnrB;
 199             j_coord_offsetC  = DIM*jnrC;
 200             j_coord_offsetD  = DIM*jnrD;
 201
 202             /* load j atom coordinates */
 203             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 204                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 205                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 206
 207             /* Calculate displacement vector */
 208             dx11             = _mm_sub_ps(ix1,jx1);
 209             dy11             = _mm_sub_ps(iy1,jy1);
 210             dz11             = _mm_sub_ps(iz1,jz1);
 211             dx12             = _mm_sub_ps(ix1,jx2);
 212             dy12             = _mm_sub_ps(iy1,jy2);
 213             dz12             = _mm_sub_ps(iz1,jz2);
 214             dx13             = _mm_sub_ps(ix1,jx3);
 215             dy13             = _mm_sub_ps(iy1,jy3);
 216             dz13             = _mm_sub_ps(iz1,jz3);
 217             dx21             = _mm_sub_ps(ix2,jx1);
 218             dy21             = _mm_sub_ps(iy2,jy1);
 219             dz21             = _mm_sub_ps(iz2,jz1);
 220             dx22             = _mm_sub_ps(ix2,jx2);
 221             dy22             = _mm_sub_ps(iy2,jy2);
 222             dz22             = _mm_sub_ps(iz2,jz2);
 223             dx23             = _mm_sub_ps(ix2,jx3);
 224             dy23             = _mm_sub_ps(iy2,jy3);
 225             dz23             = _mm_sub_ps(iz2,jz3);
 226             dx31             = _mm_sub_ps(ix3,jx1);
 227             dy31             = _mm_sub_ps(iy3,jy1);
 228             dz31             = _mm_sub_ps(iz3,jz1);
 229             dx32             = _mm_sub_ps(ix3,jx2);
 230             dy32             = _mm_sub_ps(iy3,jy2);
 231             dz32             = _mm_sub_ps(iz3,jz2);
 232             dx33             = _mm_sub_ps(ix3,jx3);
 233             dy33             = _mm_sub_ps(iy3,jy3);
 234             dz33             = _mm_sub_ps(iz3,jz3);
 235
 236             /* Calculate squared distance and things based on it */
 237             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 238             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 239             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 240             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 241             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 242             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 243             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 244             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 245             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 246
 247             rinv11           = sse2_invsqrt_f(rsq11);
 248             rinv12           = sse2_invsqrt_f(rsq12);
 249             rinv13           = sse2_invsqrt_f(rsq13);
 250             rinv21           = sse2_invsqrt_f(rsq21);
 251             rinv22           = sse2_invsqrt_f(rsq22);
 252             rinv23           = sse2_invsqrt_f(rsq23);
 253             rinv31           = sse2_invsqrt_f(rsq31);
 254             rinv32           = sse2_invsqrt_f(rsq32);
 255             rinv33           = sse2_invsqrt_f(rsq33);
 256
 257             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 258             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 259             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 260             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 261             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 262             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 263             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 264             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 265             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 266
 267             fjx1             = _mm_setzero_ps();
 268             fjy1             = _mm_setzero_ps();
 269             fjz1             = _mm_setzero_ps();
 270             fjx2             = _mm_setzero_ps();
 271             fjy2             = _mm_setzero_ps();
 272             fjz2             = _mm_setzero_ps();
 273             fjx3             = _mm_setzero_ps();
 274             fjy3             = _mm_setzero_ps();
 275             fjz3             = _mm_setzero_ps();
 276
 277             /**************************
 278              * CALCULATE INTERACTIONS *
 279              **************************/
 280
 281             /* COULOMB ELECTROSTATICS */
 282             velec            = _mm_mul_ps(qq11,rinv11);
 283             felec            = _mm_mul_ps(velec,rinvsq11);
 284
 285             /* Update potential sum for this i atom from the interaction with this j atom. */
 286             velecsum         = _mm_add_ps(velecsum,velec);
 287
 288             fscal            = felec;
 289
 290             /* Calculate temporary vectorial force */
 291             tx               = _mm_mul_ps(fscal,dx11);
 292             ty               = _mm_mul_ps(fscal,dy11);
 293             tz               = _mm_mul_ps(fscal,dz11);
 294
 295             /* Update vectorial force */
 296             fix1             = _mm_add_ps(fix1,tx);
 297             fiy1             = _mm_add_ps(fiy1,ty);
 298             fiz1             = _mm_add_ps(fiz1,tz);
 299
 300             fjx1             = _mm_add_ps(fjx1,tx);
 301             fjy1             = _mm_add_ps(fjy1,ty);
 302             fjz1             = _mm_add_ps(fjz1,tz);
 303
 304             /**************************
 305              * CALCULATE INTERACTIONS *
 306              **************************/
 307
 308             /* COULOMB ELECTROSTATICS */
 309             velec            = _mm_mul_ps(qq12,rinv12);
 310             felec            = _mm_mul_ps(velec,rinvsq12);
 311
 312             /* Update potential sum for this i atom from the interaction with this j atom. */
 313             velecsum         = _mm_add_ps(velecsum,velec);
 314
 315             fscal            = felec;
 316
 317             /* Calculate temporary vectorial force */
 318             tx               = _mm_mul_ps(fscal,dx12);
 319             ty               = _mm_mul_ps(fscal,dy12);
 320             tz               = _mm_mul_ps(fscal,dz12);
 321
 322             /* Update vectorial force */
 323             fix1             = _mm_add_ps(fix1,tx);
 324             fiy1             = _mm_add_ps(fiy1,ty);
 325             fiz1             = _mm_add_ps(fiz1,tz);
 326
 327             fjx2             = _mm_add_ps(fjx2,tx);
 328             fjy2             = _mm_add_ps(fjy2,ty);
 329             fjz2             = _mm_add_ps(fjz2,tz);
 330
 331             /**************************
 332              * CALCULATE INTERACTIONS *
 333              **************************/
 334
 335             /* COULOMB ELECTROSTATICS */
 336             velec            = _mm_mul_ps(qq13,rinv13);
 337             felec            = _mm_mul_ps(velec,rinvsq13);
 338
 339             /* Update potential sum for this i atom from the interaction with this j atom. */
 340             velecsum         = _mm_add_ps(velecsum,velec);
 341
 342             fscal            = felec;
 343
 344             /* Calculate temporary vectorial force */
 345             tx               = _mm_mul_ps(fscal,dx13);
 346             ty               = _mm_mul_ps(fscal,dy13);
 347             tz               = _mm_mul_ps(fscal,dz13);
 348
 349             /* Update vectorial force */
 350             fix1             = _mm_add_ps(fix1,tx);
 351             fiy1             = _mm_add_ps(fiy1,ty);
 352             fiz1             = _mm_add_ps(fiz1,tz);
 353
 354             fjx3             = _mm_add_ps(fjx3,tx);
 355             fjy3             = _mm_add_ps(fjy3,ty);
 356             fjz3             = _mm_add_ps(fjz3,tz);
 357
 358             /**************************
 359              * CALCULATE INTERACTIONS *
 360              **************************/
 361
 362             /* COULOMB ELECTROSTATICS */
 363             velec            = _mm_mul_ps(qq21,rinv21);
 364             felec            = _mm_mul_ps(velec,rinvsq21);
 365
 366             /* Update potential sum for this i atom from the interaction with this j atom. */
 367             velecsum         = _mm_add_ps(velecsum,velec);
 368
 369             fscal            = felec;
 370
 371             /* Calculate temporary vectorial force */
 372             tx               = _mm_mul_ps(fscal,dx21);
 373             ty               = _mm_mul_ps(fscal,dy21);
 374             tz               = _mm_mul_ps(fscal,dz21);
 375
 376             /* Update vectorial force */
 377             fix2             = _mm_add_ps(fix2,tx);
 378             fiy2             = _mm_add_ps(fiy2,ty);
 379             fiz2             = _mm_add_ps(fiz2,tz);
 380
 381             fjx1             = _mm_add_ps(fjx1,tx);
 382             fjy1             = _mm_add_ps(fjy1,ty);
 383             fjz1             = _mm_add_ps(fjz1,tz);
 384
 385             /**************************
 386              * CALCULATE INTERACTIONS *
 387              **************************/
 388
 389             /* COULOMB ELECTROSTATICS */
 390             velec            = _mm_mul_ps(qq22,rinv22);
 391             felec            = _mm_mul_ps(velec,rinvsq22);
 392
 393             /* Update potential sum for this i atom from the interaction with this j atom. */
 394             velecsum         = _mm_add_ps(velecsum,velec);
 395
 396             fscal            = felec;
 397
 398             /* Calculate temporary vectorial force */
 399             tx               = _mm_mul_ps(fscal,dx22);
 400             ty               = _mm_mul_ps(fscal,dy22);
 401             tz               = _mm_mul_ps(fscal,dz22);
 402
 403             /* Update vectorial force */
 404             fix2             = _mm_add_ps(fix2,tx);
 405             fiy2             = _mm_add_ps(fiy2,ty);
 406             fiz2             = _mm_add_ps(fiz2,tz);
 407
 408             fjx2             = _mm_add_ps(fjx2,tx);
 409             fjy2             = _mm_add_ps(fjy2,ty);
 410             fjz2             = _mm_add_ps(fjz2,tz);
 411
 412             /**************************
 413              * CALCULATE INTERACTIONS *
 414              **************************/
 415
 416             /* COULOMB ELECTROSTATICS */
 417             velec            = _mm_mul_ps(qq23,rinv23);
 418             felec            = _mm_mul_ps(velec,rinvsq23);
 419
 420             /* Update potential sum for this i atom from the interaction with this j atom. */
 421             velecsum         = _mm_add_ps(velecsum,velec);
 422
 423             fscal            = felec;
 424
 425             /* Calculate temporary vectorial force */
 426             tx               = _mm_mul_ps(fscal,dx23);
 427             ty               = _mm_mul_ps(fscal,dy23);
 428             tz               = _mm_mul_ps(fscal,dz23);
 429
 430             /* Update vectorial force */
 431             fix2             = _mm_add_ps(fix2,tx);
 432             fiy2             = _mm_add_ps(fiy2,ty);
 433             fiz2             = _mm_add_ps(fiz2,tz);
 434
 435             fjx3             = _mm_add_ps(fjx3,tx);
 436             fjy3             = _mm_add_ps(fjy3,ty);
 437             fjz3             = _mm_add_ps(fjz3,tz);
 438
 439             /**************************
 440              * CALCULATE INTERACTIONS *
 441              **************************/
 442
 443             /* COULOMB ELECTROSTATICS */
 444             velec            = _mm_mul_ps(qq31,rinv31);
 445             felec            = _mm_mul_ps(velec,rinvsq31);
 446
 447             /* Update potential sum for this i atom from the interaction with this j atom. */
 448             velecsum         = _mm_add_ps(velecsum,velec);
 449
 450             fscal            = felec;
 451
 452             /* Calculate temporary vectorial force */
 453             tx               = _mm_mul_ps(fscal,dx31);
 454             ty               = _mm_mul_ps(fscal,dy31);
 455             tz               = _mm_mul_ps(fscal,dz31);
 456
 457             /* Update vectorial force */
 458             fix3             = _mm_add_ps(fix3,tx);
 459             fiy3             = _mm_add_ps(fiy3,ty);
 460             fiz3             = _mm_add_ps(fiz3,tz);
 461
 462             fjx1             = _mm_add_ps(fjx1,tx);
 463             fjy1             = _mm_add_ps(fjy1,ty);
 464             fjz1             = _mm_add_ps(fjz1,tz);
 465
 466             /**************************
 467              * CALCULATE INTERACTIONS *
 468              **************************/
 469
 470             /* COULOMB ELECTROSTATICS */
 471             velec            = _mm_mul_ps(qq32,rinv32);
 472             felec            = _mm_mul_ps(velec,rinvsq32);
 473
 474             /* Update potential sum for this i atom from the interaction with this j atom. */
 475             velecsum         = _mm_add_ps(velecsum,velec);
 476
 477             fscal            = felec;
 478
 479             /* Calculate temporary vectorial force */
 480             tx               = _mm_mul_ps(fscal,dx32);
 481             ty               = _mm_mul_ps(fscal,dy32);
 482             tz               = _mm_mul_ps(fscal,dz32);
 483
 484             /* Update vectorial force */
 485             fix3             = _mm_add_ps(fix3,tx);
 486             fiy3             = _mm_add_ps(fiy3,ty);
 487             fiz3             = _mm_add_ps(fiz3,tz);
 488
 489             fjx2             = _mm_add_ps(fjx2,tx);
 490             fjy2             = _mm_add_ps(fjy2,ty);
 491             fjz2             = _mm_add_ps(fjz2,tz);
 492
 493             /**************************
 494              * CALCULATE INTERACTIONS *
 495              **************************/
 496
 497             /* COULOMB ELECTROSTATICS */
 498             velec            = _mm_mul_ps(qq33,rinv33);
 499             felec            = _mm_mul_ps(velec,rinvsq33);
 500
 501             /* Update potential sum for this i atom from the interaction with this j atom. */
 502             velecsum         = _mm_add_ps(velecsum,velec);
 503
 504             fscal            = felec;
 505
 506             /* Calculate temporary vectorial force */
 507             tx               = _mm_mul_ps(fscal,dx33);
 508             ty               = _mm_mul_ps(fscal,dy33);
 509             tz               = _mm_mul_ps(fscal,dz33);
 510
 511             /* Update vectorial force */
 512             fix3             = _mm_add_ps(fix3,tx);
 513             fiy3             = _mm_add_ps(fiy3,ty);
 514             fiz3             = _mm_add_ps(fiz3,tz);
 515
 516             fjx3             = _mm_add_ps(fjx3,tx);
 517             fjy3             = _mm_add_ps(fjy3,ty);
 518             fjz3             = _mm_add_ps(fjz3,tz);
 519
 520             fjptrA             = f+j_coord_offsetA;
 521             fjptrB             = f+j_coord_offsetB;
 522             fjptrC             = f+j_coord_offsetC;
 523             fjptrD             = f+j_coord_offsetD;
 524
 525             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
 526                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 527
 528             /* Inner loop uses 252 flops */
 529         }
 530
 531         if(jidx<j_index_end)
 532         {
 533
 534             /* Get j neighbor index, and coordinate index */
 535             jnrlistA         = jjnr[jidx];
 536             jnrlistB         = jjnr[jidx+1];
 537             jnrlistC         = jjnr[jidx+2];
 538             jnrlistD         = jjnr[jidx+3];
 539             /* Sign of each element will be negative for non-real atoms.
 540              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 541              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 542              */
 543             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 544             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 545             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 546             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 547             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 548             j_coord_offsetA  = DIM*jnrA;
 549             j_coord_offsetB  = DIM*jnrB;
 550             j_coord_offsetC  = DIM*jnrC;
 551             j_coord_offsetD  = DIM*jnrD;
 552
 553             /* load j atom coordinates */
 554             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 555                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 556                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 557
 558             /* Calculate displacement vector */
 559             dx11             = _mm_sub_ps(ix1,jx1);
 560             dy11             = _mm_sub_ps(iy1,jy1);
 561             dz11             = _mm_sub_ps(iz1,jz1);
 562             dx12             = _mm_sub_ps(ix1,jx2);
 563             dy12             = _mm_sub_ps(iy1,jy2);
 564             dz12             = _mm_sub_ps(iz1,jz2);
 565             dx13             = _mm_sub_ps(ix1,jx3);
 566             dy13             = _mm_sub_ps(iy1,jy3);
 567             dz13             = _mm_sub_ps(iz1,jz3);
 568             dx21             = _mm_sub_ps(ix2,jx1);
 569             dy21             = _mm_sub_ps(iy2,jy1);
 570             dz21             = _mm_sub_ps(iz2,jz1);
 571             dx22             = _mm_sub_ps(ix2,jx2);
 572             dy22             = _mm_sub_ps(iy2,jy2);
 573             dz22             = _mm_sub_ps(iz2,jz2);
 574             dx23             = _mm_sub_ps(ix2,jx3);
 575             dy23             = _mm_sub_ps(iy2,jy3);
 576             dz23             = _mm_sub_ps(iz2,jz3);
 577             dx31             = _mm_sub_ps(ix3,jx1);
 578             dy31             = _mm_sub_ps(iy3,jy1);
 579             dz31             = _mm_sub_ps(iz3,jz1);
 580             dx32             = _mm_sub_ps(ix3,jx2);
 581             dy32             = _mm_sub_ps(iy3,jy2);
 582             dz32             = _mm_sub_ps(iz3,jz2);
 583             dx33             = _mm_sub_ps(ix3,jx3);
 584             dy33             = _mm_sub_ps(iy3,jy3);
 585             dz33             = _mm_sub_ps(iz3,jz3);
 586
 587             /* Calculate squared distance and things based on it */
 588             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 589             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 590             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 591             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 592             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 593             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 594             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 595             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 596             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 597
 598             rinv11           = sse2_invsqrt_f(rsq11);
 599             rinv12           = sse2_invsqrt_f(rsq12);
 600             rinv13           = sse2_invsqrt_f(rsq13);
 601             rinv21           = sse2_invsqrt_f(rsq21);
 602             rinv22           = sse2_invsqrt_f(rsq22);
 603             rinv23           = sse2_invsqrt_f(rsq23);
 604             rinv31           = sse2_invsqrt_f(rsq31);
 605             rinv32           = sse2_invsqrt_f(rsq32);
 606             rinv33           = sse2_invsqrt_f(rsq33);
 607
 608             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 609             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 610             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 611             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 612             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 613             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 614             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 615             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 616             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 617
 618             fjx1             = _mm_setzero_ps();
 619             fjy1             = _mm_setzero_ps();
 620             fjz1             = _mm_setzero_ps();
 621             fjx2             = _mm_setzero_ps();
 622             fjy2             = _mm_setzero_ps();
 623             fjz2             = _mm_setzero_ps();
 624             fjx3             = _mm_setzero_ps();
 625             fjy3             = _mm_setzero_ps();
 626             fjz3             = _mm_setzero_ps();
 627
 628             /**************************
 629              * CALCULATE INTERACTIONS *
 630              **************************/
 631
 632             /* COULOMB ELECTROSTATICS */
 633             velec            = _mm_mul_ps(qq11,rinv11);
 634             felec            = _mm_mul_ps(velec,rinvsq11);
 635
 636             /* Update potential sum for this i atom from the interaction with this j atom. */
 637             velec            = _mm_andnot_ps(dummy_mask,velec);
 638             velecsum         = _mm_add_ps(velecsum,velec);
 639
 640             fscal            = felec;
 641
 642             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 643
 644             /* Calculate temporary vectorial force */
 645             tx               = _mm_mul_ps(fscal,dx11);
 646             ty               = _mm_mul_ps(fscal,dy11);
 647             tz               = _mm_mul_ps(fscal,dz11);
 648
 649             /* Update vectorial force */
 650             fix1             = _mm_add_ps(fix1,tx);
 651             fiy1             = _mm_add_ps(fiy1,ty);
 652             fiz1             = _mm_add_ps(fiz1,tz);
 653
 654             fjx1             = _mm_add_ps(fjx1,tx);
 655             fjy1             = _mm_add_ps(fjy1,ty);
 656             fjz1             = _mm_add_ps(fjz1,tz);
 657
 658             /**************************
 659              * CALCULATE INTERACTIONS *
 660              **************************/
 661
 662             /* COULOMB ELECTROSTATICS */
 663             velec            = _mm_mul_ps(qq12,rinv12);
 664             felec            = _mm_mul_ps(velec,rinvsq12);
 665
 666             /* Update potential sum for this i atom from the interaction with this j atom. */
 667             velec            = _mm_andnot_ps(dummy_mask,velec);
 668             velecsum         = _mm_add_ps(velecsum,velec);
 669
 670             fscal            = felec;
 671
 672             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 673
 674             /* Calculate temporary vectorial force */
 675             tx               = _mm_mul_ps(fscal,dx12);
 676             ty               = _mm_mul_ps(fscal,dy12);
 677             tz               = _mm_mul_ps(fscal,dz12);
 678
 679             /* Update vectorial force */
 680             fix1             = _mm_add_ps(fix1,tx);
 681             fiy1             = _mm_add_ps(fiy1,ty);
 682             fiz1             = _mm_add_ps(fiz1,tz);
 683
 684             fjx2             = _mm_add_ps(fjx2,tx);
 685             fjy2             = _mm_add_ps(fjy2,ty);
 686             fjz2             = _mm_add_ps(fjz2,tz);
 687
 688             /**************************
 689              * CALCULATE INTERACTIONS *
 690              **************************/
 691
 692             /* COULOMB ELECTROSTATICS */
 693             velec            = _mm_mul_ps(qq13,rinv13);
 694             felec            = _mm_mul_ps(velec,rinvsq13);
 695
 696             /* Update potential sum for this i atom from the interaction with this j atom. */
 697             velec            = _mm_andnot_ps(dummy_mask,velec);
 698             velecsum         = _mm_add_ps(velecsum,velec);
 699
 700             fscal            = felec;
 701
 702             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 703
 704             /* Calculate temporary vectorial force */
 705             tx               = _mm_mul_ps(fscal,dx13);
 706             ty               = _mm_mul_ps(fscal,dy13);
 707             tz               = _mm_mul_ps(fscal,dz13);
 708
 709             /* Update vectorial force */
 710             fix1             = _mm_add_ps(fix1,tx);
 711             fiy1             = _mm_add_ps(fiy1,ty);
 712             fiz1             = _mm_add_ps(fiz1,tz);
 713
 714             fjx3             = _mm_add_ps(fjx3,tx);
 715             fjy3             = _mm_add_ps(fjy3,ty);
 716             fjz3             = _mm_add_ps(fjz3,tz);
 717
 718             /**************************
 719              * CALCULATE INTERACTIONS *
 720              **************************/
 721
 722             /* COULOMB ELECTROSTATICS */
 723             velec            = _mm_mul_ps(qq21,rinv21);
 724             felec            = _mm_mul_ps(velec,rinvsq21);
 725
 726             /* Update potential sum for this i atom from the interaction with this j atom. */
 727             velec            = _mm_andnot_ps(dummy_mask,velec);
 728             velecsum         = _mm_add_ps(velecsum,velec);
 729
 730             fscal            = felec;
 731
 732             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 733
 734             /* Calculate temporary vectorial force */
 735             tx               = _mm_mul_ps(fscal,dx21);
 736             ty               = _mm_mul_ps(fscal,dy21);
 737             tz               = _mm_mul_ps(fscal,dz21);
 738
 739             /* Update vectorial force */
 740             fix2             = _mm_add_ps(fix2,tx);
 741             fiy2             = _mm_add_ps(fiy2,ty);
 742             fiz2             = _mm_add_ps(fiz2,tz);
 743
 744             fjx1             = _mm_add_ps(fjx1,tx);
 745             fjy1             = _mm_add_ps(fjy1,ty);
 746             fjz1             = _mm_add_ps(fjz1,tz);
 747
 748             /**************************
 749              * CALCULATE INTERACTIONS *
 750              **************************/
 751
 752             /* COULOMB ELECTROSTATICS */
 753             velec            = _mm_mul_ps(qq22,rinv22);
 754             felec            = _mm_mul_ps(velec,rinvsq22);
 755
 756             /* Update potential sum for this i atom from the interaction with this j atom. */
 757             velec            = _mm_andnot_ps(dummy_mask,velec);
 758             velecsum         = _mm_add_ps(velecsum,velec);
 759
 760             fscal            = felec;
 761
 762             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 763
 764             /* Calculate temporary vectorial force */
 765             tx               = _mm_mul_ps(fscal,dx22);
 766             ty               = _mm_mul_ps(fscal,dy22);
 767             tz               = _mm_mul_ps(fscal,dz22);
 768
 769             /* Update vectorial force */
 770             fix2             = _mm_add_ps(fix2,tx);
 771             fiy2             = _mm_add_ps(fiy2,ty);
 772             fiz2             = _mm_add_ps(fiz2,tz);
 773
 774             fjx2             = _mm_add_ps(fjx2,tx);
 775             fjy2             = _mm_add_ps(fjy2,ty);
 776             fjz2             = _mm_add_ps(fjz2,tz);
 777
 778             /**************************
 779              * CALCULATE INTERACTIONS *
 780              **************************/
 781
 782             /* COULOMB ELECTROSTATICS */
 783             velec            = _mm_mul_ps(qq23,rinv23);
 784             felec            = _mm_mul_ps(velec,rinvsq23);
 785
 786             /* Update potential sum for this i atom from the interaction with this j atom. */
 787             velec            = _mm_andnot_ps(dummy_mask,velec);
 788             velecsum         = _mm_add_ps(velecsum,velec);
 789
 790             fscal            = felec;
 791
 792             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 793
 794             /* Calculate temporary vectorial force */
 795             tx               = _mm_mul_ps(fscal,dx23);
 796             ty               = _mm_mul_ps(fscal,dy23);
 797             tz               = _mm_mul_ps(fscal,dz23);
 798
 799             /* Update vectorial force */
 800             fix2             = _mm_add_ps(fix2,tx);
 801             fiy2             = _mm_add_ps(fiy2,ty);
 802             fiz2             = _mm_add_ps(fiz2,tz);
 803
 804             fjx3             = _mm_add_ps(fjx3,tx);
 805             fjy3             = _mm_add_ps(fjy3,ty);
 806             fjz3             = _mm_add_ps(fjz3,tz);
 807
 808             /**************************
 809              * CALCULATE INTERACTIONS *
 810              **************************/
 811
 812             /* COULOMB ELECTROSTATICS */
 813             velec            = _mm_mul_ps(qq31,rinv31);
 814             felec            = _mm_mul_ps(velec,rinvsq31);
 815
 816             /* Update potential sum for this i atom from the interaction with this j atom. */
 817             velec            = _mm_andnot_ps(dummy_mask,velec);
 818             velecsum         = _mm_add_ps(velecsum,velec);
 819
 820             fscal            = felec;
 821
 822             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 823
 824             /* Calculate temporary vectorial force */
 825             tx               = _mm_mul_ps(fscal,dx31);
 826             ty               = _mm_mul_ps(fscal,dy31);
 827             tz               = _mm_mul_ps(fscal,dz31);
 828
 829             /* Update vectorial force */
 830             fix3             = _mm_add_ps(fix3,tx);
 831             fiy3             = _mm_add_ps(fiy3,ty);
 832             fiz3             = _mm_add_ps(fiz3,tz);
 833
 834             fjx1             = _mm_add_ps(fjx1,tx);
 835             fjy1             = _mm_add_ps(fjy1,ty);
 836             fjz1             = _mm_add_ps(fjz1,tz);
 837
 838             /**************************
 839              * CALCULATE INTERACTIONS *
 840              **************************/
 841
 842             /* COULOMB ELECTROSTATICS */
 843             velec            = _mm_mul_ps(qq32,rinv32);
 844             felec            = _mm_mul_ps(velec,rinvsq32);
 845
 846             /* Update potential sum for this i atom from the interaction with this j atom. */
 847             velec            = _mm_andnot_ps(dummy_mask,velec);
 848             velecsum         = _mm_add_ps(velecsum,velec);
 849
 850             fscal            = felec;
 851
 852             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 853
 854             /* Calculate temporary vectorial force */
 855             tx               = _mm_mul_ps(fscal,dx32);
 856             ty               = _mm_mul_ps(fscal,dy32);
 857             tz               = _mm_mul_ps(fscal,dz32);
 858
 859             /* Update vectorial force */
 860             fix3             = _mm_add_ps(fix3,tx);
 861             fiy3             = _mm_add_ps(fiy3,ty);
 862             fiz3             = _mm_add_ps(fiz3,tz);
 863
 864             fjx2             = _mm_add_ps(fjx2,tx);
 865             fjy2             = _mm_add_ps(fjy2,ty);
 866             fjz2             = _mm_add_ps(fjz2,tz);
 867
 868             /**************************
 869              * CALCULATE INTERACTIONS *
 870              **************************/
 871
 872             /* COULOMB ELECTROSTATICS */
 873             velec            = _mm_mul_ps(qq33,rinv33);
 874             felec            = _mm_mul_ps(velec,rinvsq33);
 875
 876             /* Update potential sum for this i atom from the interaction with this j atom. */
 877             velec            = _mm_andnot_ps(dummy_mask,velec);
 878             velecsum         = _mm_add_ps(velecsum,velec);
 879
 880             fscal            = felec;
 881
 882             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 883
 884             /* Calculate temporary vectorial force */
 885             tx               = _mm_mul_ps(fscal,dx33);
 886             ty               = _mm_mul_ps(fscal,dy33);
 887             tz               = _mm_mul_ps(fscal,dz33);
 888
 889             /* Update vectorial force */
 890             fix3             = _mm_add_ps(fix3,tx);
 891             fiy3             = _mm_add_ps(fiy3,ty);
 892             fiz3             = _mm_add_ps(fiz3,tz);
 893
 894             fjx3             = _mm_add_ps(fjx3,tx);
 895             fjy3             = _mm_add_ps(fjy3,ty);
 896             fjz3             = _mm_add_ps(fjz3,tz);
 897
 898             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 899             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 900             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 901             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 902
 903             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
 904                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 905
 906             /* Inner loop uses 252 flops */
 907         }
 908
 909         /* End of innermost loop */
 910
 911         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 912                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
 913
 914         ggid                        = gid[iidx];
 915         /* Update potential energies */
 916         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 917
 918         /* Increment number of inner iterations */
 919         inneriter                  += j_index_end - j_index_start;
 920
 921         /* Outer loop uses 19 flops */
 922     }
 923
 924     /* Increment number of outer iterations */
 925     outeriter        += nri;
 926
 927     /* Update outer/inner flops */
 928
 929     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*252);
 930 }
 931 /*
 932  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwNone_GeomW4W4_F_sse2_single
 933  * Electrostatics interaction: Coulomb
 934  * VdW interaction:            None
 935  * Geometry:                   Water4-Water4
 936  * Calculate force/pot:        Force
 937  */
 938 void
 939 nb_kernel_ElecCoul_VdwNone_GeomW4W4_F_sse2_single
 940                     (t_nblist                    * gmx_restrict       nlist,
 941                      rvec                        * gmx_restrict          xx,
 942                      rvec                        * gmx_restrict          ff,
 943                      struct t_forcerec           * gmx_restrict          fr,
 944                      t_mdatoms                   * gmx_restrict     mdatoms,
 945                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 946                      t_nrnb                      * gmx_restrict        nrnb)
 947 {
 948     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 949      * just 0 for non-waters.
 950      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 951      * jnr indices corresponding to data put in the four positions in the SIMD register.
 952      */
 953     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 954     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 955     int              jnrA,jnrB,jnrC,jnrD;
 956     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 957     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 958     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 959     real             rcutoff_scalar;
 960     real             *shiftvec,*fshift,*x,*f;
 961     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 962     real             scratch[4*DIM];
 963     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 964     int              vdwioffset1;
 965     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 966     int              vdwioffset2;
 967     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 968     int              vdwioffset3;
 969     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 970     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
 971     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 972     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
 973     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 974     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
 975     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 976     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 977     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 978     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 979     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 980     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 981     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 982     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 983     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 984     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 985     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 986     real             *charge;
 987     __m128           dummy_mask,cutoff_mask;
 988     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 989     __m128           one     = _mm_set1_ps(1.0);
 990     __m128           two     = _mm_set1_ps(2.0);
 991     x                = xx[0];
 992     f                = ff[0];
 993
 994     nri              = nlist->nri;
 995     iinr             = nlist->iinr;
 996     jindex           = nlist->jindex;
 997     jjnr             = nlist->jjnr;
 998     shiftidx         = nlist->shift;
 999     gid              = nlist->gid;
1000     shiftvec         = fr->shift_vec[0];
1001     fshift           = fr->fshift[0];
1002     facel            = _mm_set1_ps(fr->ic->epsfac);
1003     charge           = mdatoms->chargeA;
1004
1005     /* Setup water-specific parameters */
1006     inr              = nlist->iinr[0];
1007     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1008     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1009     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1010
1011     jq1              = _mm_set1_ps(charge[inr+1]);
1012     jq2              = _mm_set1_ps(charge[inr+2]);
1013     jq3              = _mm_set1_ps(charge[inr+3]);
1014     qq11             = _mm_mul_ps(iq1,jq1);
1015     qq12             = _mm_mul_ps(iq1,jq2);
1016     qq13             = _mm_mul_ps(iq1,jq3);
1017     qq21             = _mm_mul_ps(iq2,jq1);
1018     qq22             = _mm_mul_ps(iq2,jq2);
1019     qq23             = _mm_mul_ps(iq2,jq3);
1020     qq31             = _mm_mul_ps(iq3,jq1);
1021     qq32             = _mm_mul_ps(iq3,jq2);
1022     qq33             = _mm_mul_ps(iq3,jq3);
1023
1024     /* Avoid stupid compiler warnings */
1025     jnrA = jnrB = jnrC = jnrD = 0;
1026     j_coord_offsetA = 0;
1027     j_coord_offsetB = 0;
1028     j_coord_offsetC = 0;
1029     j_coord_offsetD = 0;
1030
1031     outeriter        = 0;
1032     inneriter        = 0;
1033
1034     for(iidx=0;iidx<4*DIM;iidx++)
1035     {
1036         scratch[iidx] = 0.0;
1037     }
1038
1039     /* Start outer loop over neighborlists */
1040     for(iidx=0; iidx<nri; iidx++)
1041     {
1042         /* Load shift vector for this list */
1043         i_shift_offset   = DIM*shiftidx[iidx];
1044
1045         /* Load limits for loop over neighbors */
1046         j_index_start    = jindex[iidx];
1047         j_index_end      = jindex[iidx+1];
1048
1049         /* Get outer coordinate index */
1050         inr              = iinr[iidx];
1051         i_coord_offset   = DIM*inr;
1052
1053         /* Load i particle coords and add shift vector */
1054         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
1055                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1056
1057         fix1             = _mm_setzero_ps();
1058         fiy1             = _mm_setzero_ps();
1059         fiz1             = _mm_setzero_ps();
1060         fix2             = _mm_setzero_ps();
1061         fiy2             = _mm_setzero_ps();
1062         fiz2             = _mm_setzero_ps();
1063         fix3             = _mm_setzero_ps();
1064         fiy3             = _mm_setzero_ps();
1065         fiz3             = _mm_setzero_ps();
1066
1067         /* Start inner kernel loop */
1068         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1069         {
1070
1071             /* Get j neighbor index, and coordinate index */
1072             jnrA             = jjnr[jidx];
1073             jnrB             = jjnr[jidx+1];
1074             jnrC             = jjnr[jidx+2];
1075             jnrD             = jjnr[jidx+3];
1076             j_coord_offsetA  = DIM*jnrA;
1077             j_coord_offsetB  = DIM*jnrB;
1078             j_coord_offsetC  = DIM*jnrC;
1079             j_coord_offsetD  = DIM*jnrD;
1080
1081             /* load j atom coordinates */
1082             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1083                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1084                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1085
1086             /* Calculate displacement vector */
1087             dx11             = _mm_sub_ps(ix1,jx1);
1088             dy11             = _mm_sub_ps(iy1,jy1);
1089             dz11             = _mm_sub_ps(iz1,jz1);
1090             dx12             = _mm_sub_ps(ix1,jx2);
1091             dy12             = _mm_sub_ps(iy1,jy2);
1092             dz12             = _mm_sub_ps(iz1,jz2);
1093             dx13             = _mm_sub_ps(ix1,jx3);
1094             dy13             = _mm_sub_ps(iy1,jy3);
1095             dz13             = _mm_sub_ps(iz1,jz3);
1096             dx21             = _mm_sub_ps(ix2,jx1);
1097             dy21             = _mm_sub_ps(iy2,jy1);
1098             dz21             = _mm_sub_ps(iz2,jz1);
1099             dx22             = _mm_sub_ps(ix2,jx2);
1100             dy22             = _mm_sub_ps(iy2,jy2);
1101             dz22             = _mm_sub_ps(iz2,jz2);
1102             dx23             = _mm_sub_ps(ix2,jx3);
1103             dy23             = _mm_sub_ps(iy2,jy3);
1104             dz23             = _mm_sub_ps(iz2,jz3);
1105             dx31             = _mm_sub_ps(ix3,jx1);
1106             dy31             = _mm_sub_ps(iy3,jy1);
1107             dz31             = _mm_sub_ps(iz3,jz1);
1108             dx32             = _mm_sub_ps(ix3,jx2);
1109             dy32             = _mm_sub_ps(iy3,jy2);
1110             dz32             = _mm_sub_ps(iz3,jz2);
1111             dx33             = _mm_sub_ps(ix3,jx3);
1112             dy33             = _mm_sub_ps(iy3,jy3);
1113             dz33             = _mm_sub_ps(iz3,jz3);
1114
1115             /* Calculate squared distance and things based on it */
1116             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1117             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1118             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1119             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1120             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1121             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1122             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1123             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1124             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1125
1126             rinv11           = sse2_invsqrt_f(rsq11);
1127             rinv12           = sse2_invsqrt_f(rsq12);
1128             rinv13           = sse2_invsqrt_f(rsq13);
1129             rinv21           = sse2_invsqrt_f(rsq21);
1130             rinv22           = sse2_invsqrt_f(rsq22);
1131             rinv23           = sse2_invsqrt_f(rsq23);
1132             rinv31           = sse2_invsqrt_f(rsq31);
1133             rinv32           = sse2_invsqrt_f(rsq32);
1134             rinv33           = sse2_invsqrt_f(rsq33);
1135
1136             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1137             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1138             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1139             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1140             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1141             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1142             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1143             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1144             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1145
1146             fjx1             = _mm_setzero_ps();
1147             fjy1             = _mm_setzero_ps();
1148             fjz1             = _mm_setzero_ps();
1149             fjx2             = _mm_setzero_ps();
1150             fjy2             = _mm_setzero_ps();
1151             fjz2             = _mm_setzero_ps();
1152             fjx3             = _mm_setzero_ps();
1153             fjy3             = _mm_setzero_ps();
1154             fjz3             = _mm_setzero_ps();
1155
1156             /**************************
1157              * CALCULATE INTERACTIONS *
1158              **************************/
1159
1160             /* COULOMB ELECTROSTATICS */
1161             velec            = _mm_mul_ps(qq11,rinv11);
1162             felec            = _mm_mul_ps(velec,rinvsq11);
1163
1164             fscal            = felec;
1165
1166             /* Calculate temporary vectorial force */
1167             tx               = _mm_mul_ps(fscal,dx11);
1168             ty               = _mm_mul_ps(fscal,dy11);
1169             tz               = _mm_mul_ps(fscal,dz11);
1170
1171             /* Update vectorial force */
1172             fix1             = _mm_add_ps(fix1,tx);
1173             fiy1             = _mm_add_ps(fiy1,ty);
1174             fiz1             = _mm_add_ps(fiz1,tz);
1175
1176             fjx1             = _mm_add_ps(fjx1,tx);
1177             fjy1             = _mm_add_ps(fjy1,ty);
1178             fjz1             = _mm_add_ps(fjz1,tz);
1179
1180             /**************************
1181              * CALCULATE INTERACTIONS *
1182              **************************/
1183
1184             /* COULOMB ELECTROSTATICS */
1185             velec            = _mm_mul_ps(qq12,rinv12);
1186             felec            = _mm_mul_ps(velec,rinvsq12);
1187
1188             fscal            = felec;
1189
1190             /* Calculate temporary vectorial force */
1191             tx               = _mm_mul_ps(fscal,dx12);
1192             ty               = _mm_mul_ps(fscal,dy12);
1193             tz               = _mm_mul_ps(fscal,dz12);
1194
1195             /* Update vectorial force */
1196             fix1             = _mm_add_ps(fix1,tx);
1197             fiy1             = _mm_add_ps(fiy1,ty);
1198             fiz1             = _mm_add_ps(fiz1,tz);
1199
1200             fjx2             = _mm_add_ps(fjx2,tx);
1201             fjy2             = _mm_add_ps(fjy2,ty);
1202             fjz2             = _mm_add_ps(fjz2,tz);
1203
1204             /**************************
1205              * CALCULATE INTERACTIONS *
1206              **************************/
1207
1208             /* COULOMB ELECTROSTATICS */
1209             velec            = _mm_mul_ps(qq13,rinv13);
1210             felec            = _mm_mul_ps(velec,rinvsq13);
1211
1212             fscal            = felec;
1213
1214             /* Calculate temporary vectorial force */
1215             tx               = _mm_mul_ps(fscal,dx13);
1216             ty               = _mm_mul_ps(fscal,dy13);
1217             tz               = _mm_mul_ps(fscal,dz13);
1218
1219             /* Update vectorial force */
1220             fix1             = _mm_add_ps(fix1,tx);
1221             fiy1             = _mm_add_ps(fiy1,ty);
1222             fiz1             = _mm_add_ps(fiz1,tz);
1223
1224             fjx3             = _mm_add_ps(fjx3,tx);
1225             fjy3             = _mm_add_ps(fjy3,ty);
1226             fjz3             = _mm_add_ps(fjz3,tz);
1227
1228             /**************************
1229              * CALCULATE INTERACTIONS *
1230              **************************/
1231
1232             /* COULOMB ELECTROSTATICS */
1233             velec            = _mm_mul_ps(qq21,rinv21);
1234             felec            = _mm_mul_ps(velec,rinvsq21);
1235
1236             fscal            = felec;
1237
1238             /* Calculate temporary vectorial force */
1239             tx               = _mm_mul_ps(fscal,dx21);
1240             ty               = _mm_mul_ps(fscal,dy21);
1241             tz               = _mm_mul_ps(fscal,dz21);
1242
1243             /* Update vectorial force */
1244             fix2             = _mm_add_ps(fix2,tx);
1245             fiy2             = _mm_add_ps(fiy2,ty);
1246             fiz2             = _mm_add_ps(fiz2,tz);
1247
1248             fjx1             = _mm_add_ps(fjx1,tx);
1249             fjy1             = _mm_add_ps(fjy1,ty);
1250             fjz1             = _mm_add_ps(fjz1,tz);
1251
1252             /**************************
1253              * CALCULATE INTERACTIONS *
1254              **************************/
1255
1256             /* COULOMB ELECTROSTATICS */
1257             velec            = _mm_mul_ps(qq22,rinv22);
1258             felec            = _mm_mul_ps(velec,rinvsq22);
1259
1260             fscal            = felec;
1261
1262             /* Calculate temporary vectorial force */
1263             tx               = _mm_mul_ps(fscal,dx22);
1264             ty               = _mm_mul_ps(fscal,dy22);
1265             tz               = _mm_mul_ps(fscal,dz22);
1266
1267             /* Update vectorial force */
1268             fix2             = _mm_add_ps(fix2,tx);
1269             fiy2             = _mm_add_ps(fiy2,ty);
1270             fiz2             = _mm_add_ps(fiz2,tz);
1271
1272             fjx2             = _mm_add_ps(fjx2,tx);
1273             fjy2             = _mm_add_ps(fjy2,ty);
1274             fjz2             = _mm_add_ps(fjz2,tz);
1275
1276             /**************************
1277              * CALCULATE INTERACTIONS *
1278              **************************/
1279
1280             /* COULOMB ELECTROSTATICS */
1281             velec            = _mm_mul_ps(qq23,rinv23);
1282             felec            = _mm_mul_ps(velec,rinvsq23);
1283
1284             fscal            = felec;
1285
1286             /* Calculate temporary vectorial force */
1287             tx               = _mm_mul_ps(fscal,dx23);
1288             ty               = _mm_mul_ps(fscal,dy23);
1289             tz               = _mm_mul_ps(fscal,dz23);
1290
1291             /* Update vectorial force */
1292             fix2             = _mm_add_ps(fix2,tx);
1293             fiy2             = _mm_add_ps(fiy2,ty);
1294             fiz2             = _mm_add_ps(fiz2,tz);
1295
1296             fjx3             = _mm_add_ps(fjx3,tx);
1297             fjy3             = _mm_add_ps(fjy3,ty);
1298             fjz3             = _mm_add_ps(fjz3,tz);
1299
1300             /**************************
1301              * CALCULATE INTERACTIONS *
1302              **************************/
1303
1304             /* COULOMB ELECTROSTATICS */
1305             velec            = _mm_mul_ps(qq31,rinv31);
1306             felec            = _mm_mul_ps(velec,rinvsq31);
1307
1308             fscal            = felec;
1309
1310             /* Calculate temporary vectorial force */
1311             tx               = _mm_mul_ps(fscal,dx31);
1312             ty               = _mm_mul_ps(fscal,dy31);
1313             tz               = _mm_mul_ps(fscal,dz31);
1314
1315             /* Update vectorial force */
1316             fix3             = _mm_add_ps(fix3,tx);
1317             fiy3             = _mm_add_ps(fiy3,ty);
1318             fiz3             = _mm_add_ps(fiz3,tz);
1319
1320             fjx1             = _mm_add_ps(fjx1,tx);
1321             fjy1             = _mm_add_ps(fjy1,ty);
1322             fjz1             = _mm_add_ps(fjz1,tz);
1323
1324             /**************************
1325              * CALCULATE INTERACTIONS *
1326              **************************/
1327
1328             /* COULOMB ELECTROSTATICS */
1329             velec            = _mm_mul_ps(qq32,rinv32);
1330             felec            = _mm_mul_ps(velec,rinvsq32);
1331
1332             fscal            = felec;
1333
1334             /* Calculate temporary vectorial force */
1335             tx               = _mm_mul_ps(fscal,dx32);
1336             ty               = _mm_mul_ps(fscal,dy32);
1337             tz               = _mm_mul_ps(fscal,dz32);
1338
1339             /* Update vectorial force */
1340             fix3             = _mm_add_ps(fix3,tx);
1341             fiy3             = _mm_add_ps(fiy3,ty);
1342             fiz3             = _mm_add_ps(fiz3,tz);
1343
1344             fjx2             = _mm_add_ps(fjx2,tx);
1345             fjy2             = _mm_add_ps(fjy2,ty);
1346             fjz2             = _mm_add_ps(fjz2,tz);
1347
1348             /**************************
1349              * CALCULATE INTERACTIONS *
1350              **************************/
1351
1352             /* COULOMB ELECTROSTATICS */
1353             velec            = _mm_mul_ps(qq33,rinv33);
1354             felec            = _mm_mul_ps(velec,rinvsq33);
1355
1356             fscal            = felec;
1357
1358             /* Calculate temporary vectorial force */
1359             tx               = _mm_mul_ps(fscal,dx33);
1360             ty               = _mm_mul_ps(fscal,dy33);
1361             tz               = _mm_mul_ps(fscal,dz33);
1362
1363             /* Update vectorial force */
1364             fix3             = _mm_add_ps(fix3,tx);
1365             fiy3             = _mm_add_ps(fiy3,ty);
1366             fiz3             = _mm_add_ps(fiz3,tz);
1367
1368             fjx3             = _mm_add_ps(fjx3,tx);
1369             fjy3             = _mm_add_ps(fjy3,ty);
1370             fjz3             = _mm_add_ps(fjz3,tz);
1371
1372             fjptrA             = f+j_coord_offsetA;
1373             fjptrB             = f+j_coord_offsetB;
1374             fjptrC             = f+j_coord_offsetC;
1375             fjptrD             = f+j_coord_offsetD;
1376
1377             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1378                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1379
1380             /* Inner loop uses 243 flops */
1381         }
1382
1383         if(jidx<j_index_end)
1384         {
1385
1386             /* Get j neighbor index, and coordinate index */
1387             jnrlistA         = jjnr[jidx];
1388             jnrlistB         = jjnr[jidx+1];
1389             jnrlistC         = jjnr[jidx+2];
1390             jnrlistD         = jjnr[jidx+3];
1391             /* Sign of each element will be negative for non-real atoms.
1392              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1393              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1394              */
1395             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1396             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1397             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1398             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1399             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1400             j_coord_offsetA  = DIM*jnrA;
1401             j_coord_offsetB  = DIM*jnrB;
1402             j_coord_offsetC  = DIM*jnrC;
1403             j_coord_offsetD  = DIM*jnrD;
1404
1405             /* load j atom coordinates */
1406             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1407                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1408                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1409
1410             /* Calculate displacement vector */
1411             dx11             = _mm_sub_ps(ix1,jx1);
1412             dy11             = _mm_sub_ps(iy1,jy1);
1413             dz11             = _mm_sub_ps(iz1,jz1);
1414             dx12             = _mm_sub_ps(ix1,jx2);
1415             dy12             = _mm_sub_ps(iy1,jy2);
1416             dz12             = _mm_sub_ps(iz1,jz2);
1417             dx13             = _mm_sub_ps(ix1,jx3);
1418             dy13             = _mm_sub_ps(iy1,jy3);
1419             dz13             = _mm_sub_ps(iz1,jz3);
1420             dx21             = _mm_sub_ps(ix2,jx1);
1421             dy21             = _mm_sub_ps(iy2,jy1);
1422             dz21             = _mm_sub_ps(iz2,jz1);
1423             dx22             = _mm_sub_ps(ix2,jx2);
1424             dy22             = _mm_sub_ps(iy2,jy2);
1425             dz22             = _mm_sub_ps(iz2,jz2);
1426             dx23             = _mm_sub_ps(ix2,jx3);
1427             dy23             = _mm_sub_ps(iy2,jy3);
1428             dz23             = _mm_sub_ps(iz2,jz3);
1429             dx31             = _mm_sub_ps(ix3,jx1);
1430             dy31             = _mm_sub_ps(iy3,jy1);
1431             dz31             = _mm_sub_ps(iz3,jz1);
1432             dx32             = _mm_sub_ps(ix3,jx2);
1433             dy32             = _mm_sub_ps(iy3,jy2);
1434             dz32             = _mm_sub_ps(iz3,jz2);
1435             dx33             = _mm_sub_ps(ix3,jx3);
1436             dy33             = _mm_sub_ps(iy3,jy3);
1437             dz33             = _mm_sub_ps(iz3,jz3);
1438
1439             /* Calculate squared distance and things based on it */
1440             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1441             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1442             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1443             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1444             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1445             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1446             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1447             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1448             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1449
1450             rinv11           = sse2_invsqrt_f(rsq11);
1451             rinv12           = sse2_invsqrt_f(rsq12);
1452             rinv13           = sse2_invsqrt_f(rsq13);
1453             rinv21           = sse2_invsqrt_f(rsq21);
1454             rinv22           = sse2_invsqrt_f(rsq22);
1455             rinv23           = sse2_invsqrt_f(rsq23);
1456             rinv31           = sse2_invsqrt_f(rsq31);
1457             rinv32           = sse2_invsqrt_f(rsq32);
1458             rinv33           = sse2_invsqrt_f(rsq33);
1459
1460             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1461             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1462             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1463             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1464             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1465             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1466             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1467             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1468             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1469
1470             fjx1             = _mm_setzero_ps();
1471             fjy1             = _mm_setzero_ps();
1472             fjz1             = _mm_setzero_ps();
1473             fjx2             = _mm_setzero_ps();
1474             fjy2             = _mm_setzero_ps();
1475             fjz2             = _mm_setzero_ps();
1476             fjx3             = _mm_setzero_ps();
1477             fjy3             = _mm_setzero_ps();
1478             fjz3             = _mm_setzero_ps();
1479
1480             /**************************
1481              * CALCULATE INTERACTIONS *
1482              **************************/
1483
1484             /* COULOMB ELECTROSTATICS */
1485             velec            = _mm_mul_ps(qq11,rinv11);
1486             felec            = _mm_mul_ps(velec,rinvsq11);
1487
1488             fscal            = felec;
1489
1490             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1491
1492             /* Calculate temporary vectorial force */
1493             tx               = _mm_mul_ps(fscal,dx11);
1494             ty               = _mm_mul_ps(fscal,dy11);
1495             tz               = _mm_mul_ps(fscal,dz11);
1496
1497             /* Update vectorial force */
1498             fix1             = _mm_add_ps(fix1,tx);
1499             fiy1             = _mm_add_ps(fiy1,ty);
1500             fiz1             = _mm_add_ps(fiz1,tz);
1501
1502             fjx1             = _mm_add_ps(fjx1,tx);
1503             fjy1             = _mm_add_ps(fjy1,ty);
1504             fjz1             = _mm_add_ps(fjz1,tz);
1505
1506             /**************************
1507              * CALCULATE INTERACTIONS *
1508              **************************/
1509
1510             /* COULOMB ELECTROSTATICS */
1511             velec            = _mm_mul_ps(qq12,rinv12);
1512             felec            = _mm_mul_ps(velec,rinvsq12);
1513
1514             fscal            = felec;
1515
1516             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1517
1518             /* Calculate temporary vectorial force */
1519             tx               = _mm_mul_ps(fscal,dx12);
1520             ty               = _mm_mul_ps(fscal,dy12);
1521             tz               = _mm_mul_ps(fscal,dz12);
1522
1523             /* Update vectorial force */
1524             fix1             = _mm_add_ps(fix1,tx);
1525             fiy1             = _mm_add_ps(fiy1,ty);
1526             fiz1             = _mm_add_ps(fiz1,tz);
1527
1528             fjx2             = _mm_add_ps(fjx2,tx);
1529             fjy2             = _mm_add_ps(fjy2,ty);
1530             fjz2             = _mm_add_ps(fjz2,tz);
1531
1532             /**************************
1533              * CALCULATE INTERACTIONS *
1534              **************************/
1535
1536             /* COULOMB ELECTROSTATICS */
1537             velec            = _mm_mul_ps(qq13,rinv13);
1538             felec            = _mm_mul_ps(velec,rinvsq13);
1539
1540             fscal            = felec;
1541
1542             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1543
1544             /* Calculate temporary vectorial force */
1545             tx               = _mm_mul_ps(fscal,dx13);
1546             ty               = _mm_mul_ps(fscal,dy13);
1547             tz               = _mm_mul_ps(fscal,dz13);
1548
1549             /* Update vectorial force */
1550             fix1             = _mm_add_ps(fix1,tx);
1551             fiy1             = _mm_add_ps(fiy1,ty);
1552             fiz1             = _mm_add_ps(fiz1,tz);
1553
1554             fjx3             = _mm_add_ps(fjx3,tx);
1555             fjy3             = _mm_add_ps(fjy3,ty);
1556             fjz3             = _mm_add_ps(fjz3,tz);
1557
1558             /**************************
1559              * CALCULATE INTERACTIONS *
1560              **************************/
1561
1562             /* COULOMB ELECTROSTATICS */
1563             velec            = _mm_mul_ps(qq21,rinv21);
1564             felec            = _mm_mul_ps(velec,rinvsq21);
1565
1566             fscal            = felec;
1567
1568             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1569
1570             /* Calculate temporary vectorial force */
1571             tx               = _mm_mul_ps(fscal,dx21);
1572             ty               = _mm_mul_ps(fscal,dy21);
1573             tz               = _mm_mul_ps(fscal,dz21);
1574
1575             /* Update vectorial force */
1576             fix2             = _mm_add_ps(fix2,tx);
1577             fiy2             = _mm_add_ps(fiy2,ty);
1578             fiz2             = _mm_add_ps(fiz2,tz);
1579
1580             fjx1             = _mm_add_ps(fjx1,tx);
1581             fjy1             = _mm_add_ps(fjy1,ty);
1582             fjz1             = _mm_add_ps(fjz1,tz);
1583
1584             /**************************
1585              * CALCULATE INTERACTIONS *
1586              **************************/
1587
1588             /* COULOMB ELECTROSTATICS */
1589             velec            = _mm_mul_ps(qq22,rinv22);
1590             felec            = _mm_mul_ps(velec,rinvsq22);
1591
1592             fscal            = felec;
1593
1594             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1595
1596             /* Calculate temporary vectorial force */
1597             tx               = _mm_mul_ps(fscal,dx22);
1598             ty               = _mm_mul_ps(fscal,dy22);
1599             tz               = _mm_mul_ps(fscal,dz22);
1600
1601             /* Update vectorial force */
1602             fix2             = _mm_add_ps(fix2,tx);
1603             fiy2             = _mm_add_ps(fiy2,ty);
1604             fiz2             = _mm_add_ps(fiz2,tz);
1605
1606             fjx2             = _mm_add_ps(fjx2,tx);
1607             fjy2             = _mm_add_ps(fjy2,ty);
1608             fjz2             = _mm_add_ps(fjz2,tz);
1609
1610             /**************************
1611              * CALCULATE INTERACTIONS *
1612              **************************/
1613
1614             /* COULOMB ELECTROSTATICS */
1615             velec            = _mm_mul_ps(qq23,rinv23);
1616             felec            = _mm_mul_ps(velec,rinvsq23);
1617
1618             fscal            = felec;
1619
1620             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1621
1622             /* Calculate temporary vectorial force */
1623             tx               = _mm_mul_ps(fscal,dx23);
1624             ty               = _mm_mul_ps(fscal,dy23);
1625             tz               = _mm_mul_ps(fscal,dz23);
1626
1627             /* Update vectorial force */
1628             fix2             = _mm_add_ps(fix2,tx);
1629             fiy2             = _mm_add_ps(fiy2,ty);
1630             fiz2             = _mm_add_ps(fiz2,tz);
1631
1632             fjx3             = _mm_add_ps(fjx3,tx);
1633             fjy3             = _mm_add_ps(fjy3,ty);
1634             fjz3             = _mm_add_ps(fjz3,tz);
1635
1636             /**************************
1637              * CALCULATE INTERACTIONS *
1638              **************************/
1639
1640             /* COULOMB ELECTROSTATICS */
1641             velec            = _mm_mul_ps(qq31,rinv31);
1642             felec            = _mm_mul_ps(velec,rinvsq31);
1643
1644             fscal            = felec;
1645
1646             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1647
1648             /* Calculate temporary vectorial force */
1649             tx               = _mm_mul_ps(fscal,dx31);
1650             ty               = _mm_mul_ps(fscal,dy31);
1651             tz               = _mm_mul_ps(fscal,dz31);
1652
1653             /* Update vectorial force */
1654             fix3             = _mm_add_ps(fix3,tx);
1655             fiy3             = _mm_add_ps(fiy3,ty);
1656             fiz3             = _mm_add_ps(fiz3,tz);
1657
1658             fjx1             = _mm_add_ps(fjx1,tx);
1659             fjy1             = _mm_add_ps(fjy1,ty);
1660             fjz1             = _mm_add_ps(fjz1,tz);
1661
1662             /**************************
1663              * CALCULATE INTERACTIONS *
1664              **************************/
1665
1666             /* COULOMB ELECTROSTATICS */
1667             velec            = _mm_mul_ps(qq32,rinv32);
1668             felec            = _mm_mul_ps(velec,rinvsq32);
1669
1670             fscal            = felec;
1671
1672             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1673
1674             /* Calculate temporary vectorial force */
1675             tx               = _mm_mul_ps(fscal,dx32);
1676             ty               = _mm_mul_ps(fscal,dy32);
1677             tz               = _mm_mul_ps(fscal,dz32);
1678
1679             /* Update vectorial force */
1680             fix3             = _mm_add_ps(fix3,tx);
1681             fiy3             = _mm_add_ps(fiy3,ty);
1682             fiz3             = _mm_add_ps(fiz3,tz);
1683
1684             fjx2             = _mm_add_ps(fjx2,tx);
1685             fjy2             = _mm_add_ps(fjy2,ty);
1686             fjz2             = _mm_add_ps(fjz2,tz);
1687
1688             /**************************
1689              * CALCULATE INTERACTIONS *
1690              **************************/
1691
1692             /* COULOMB ELECTROSTATICS */
1693             velec            = _mm_mul_ps(qq33,rinv33);
1694             felec            = _mm_mul_ps(velec,rinvsq33);
1695
1696             fscal            = felec;
1697
1698             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1699
1700             /* Calculate temporary vectorial force */
1701             tx               = _mm_mul_ps(fscal,dx33);
1702             ty               = _mm_mul_ps(fscal,dy33);
1703             tz               = _mm_mul_ps(fscal,dz33);
1704
1705             /* Update vectorial force */
1706             fix3             = _mm_add_ps(fix3,tx);
1707             fiy3             = _mm_add_ps(fiy3,ty);
1708             fiz3             = _mm_add_ps(fiz3,tz);
1709
1710             fjx3             = _mm_add_ps(fjx3,tx);
1711             fjy3             = _mm_add_ps(fjy3,ty);
1712             fjz3             = _mm_add_ps(fjz3,tz);
1713
1714             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1715             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1716             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1717             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1718
1719             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1720                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1721
1722             /* Inner loop uses 243 flops */
1723         }
1724
1725         /* End of innermost loop */
1726
1727         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1728                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1729
1730         /* Increment number of inner iterations */
1731         inneriter                  += j_index_end - j_index_start;
1732
1733         /* Outer loop uses 18 flops */
1734     }
1735
1736     /* Increment number of outer iterations */
1737     outeriter        += nri;
1738
1739     /* Update outer/inner flops */
1740
1741     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*243);
1742 }