src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecRF_VdwNone_GeomW4W4_sse2_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS 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/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_sse2_single.h"
  50 #include "kernelutil_x86_sse2_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwNone_GeomW4W4_VF_sse2_single
  54  * Electrostatics interaction: ReactionField
  55  * VdW interaction:            None
  56  * Geometry:                   Water4-Water4
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecRF_VdwNone_GeomW4W4_VF_sse2_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              vdwioffset1;
  86     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  87     int              vdwioffset2;
  88     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  89     int              vdwioffset3;
  90     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  91     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  92     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  93     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  94     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  95     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  96     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  97     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  98     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  99     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 100     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 101     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 102     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 103     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 104     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 105     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 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     krf              = _mm_set1_ps(fr->ic->k_rf);
 126     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 127     crf              = _mm_set1_ps(fr->ic->c_rf);
 128
 129     /* Setup water-specific parameters */
 130     inr              = nlist->iinr[0];
 131     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 132     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 133     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 134
 135     jq1              = _mm_set1_ps(charge[inr+1]);
 136     jq2              = _mm_set1_ps(charge[inr+2]);
 137     jq3              = _mm_set1_ps(charge[inr+3]);
 138     qq11             = _mm_mul_ps(iq1,jq1);
 139     qq12             = _mm_mul_ps(iq1,jq2);
 140     qq13             = _mm_mul_ps(iq1,jq3);
 141     qq21             = _mm_mul_ps(iq2,jq1);
 142     qq22             = _mm_mul_ps(iq2,jq2);
 143     qq23             = _mm_mul_ps(iq2,jq3);
 144     qq31             = _mm_mul_ps(iq3,jq1);
 145     qq32             = _mm_mul_ps(iq3,jq2);
 146     qq33             = _mm_mul_ps(iq3,jq3);
 147
 148     /* Avoid stupid compiler warnings */
 149     jnrA = jnrB = jnrC = jnrD = 0;
 150     j_coord_offsetA = 0;
 151     j_coord_offsetB = 0;
 152     j_coord_offsetC = 0;
 153     j_coord_offsetD = 0;
 154
 155     outeriter        = 0;
 156     inneriter        = 0;
 157
 158     for(iidx=0;iidx<4*DIM;iidx++)
 159     {
 160         scratch[iidx] = 0.0;
 161     }
 162
 163     /* Start outer loop over neighborlists */
 164     for(iidx=0; iidx<nri; iidx++)
 165     {
 166         /* Load shift vector for this list */
 167         i_shift_offset   = DIM*shiftidx[iidx];
 168
 169         /* Load limits for loop over neighbors */
 170         j_index_start    = jindex[iidx];
 171         j_index_end      = jindex[iidx+1];
 172
 173         /* Get outer coordinate index */
 174         inr              = iinr[iidx];
 175         i_coord_offset   = DIM*inr;
 176
 177         /* Load i particle coords and add shift vector */
 178         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 179                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 180
 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         fix3             = _mm_setzero_ps();
 188         fiy3             = _mm_setzero_ps();
 189         fiz3             = _mm_setzero_ps();
 190
 191         /* Reset potential sums */
 192         velecsum         = _mm_setzero_ps();
 193
 194         /* Start inner kernel loop */
 195         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 196         {
 197
 198             /* Get j neighbor index, and coordinate index */
 199             jnrA             = jjnr[jidx];
 200             jnrB             = jjnr[jidx+1];
 201             jnrC             = jjnr[jidx+2];
 202             jnrD             = jjnr[jidx+3];
 203             j_coord_offsetA  = DIM*jnrA;
 204             j_coord_offsetB  = DIM*jnrB;
 205             j_coord_offsetC  = DIM*jnrC;
 206             j_coord_offsetD  = DIM*jnrD;
 207
 208             /* load j atom coordinates */
 209             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 210                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 211                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 212
 213             /* Calculate displacement vector */
 214             dx11             = _mm_sub_ps(ix1,jx1);
 215             dy11             = _mm_sub_ps(iy1,jy1);
 216             dz11             = _mm_sub_ps(iz1,jz1);
 217             dx12             = _mm_sub_ps(ix1,jx2);
 218             dy12             = _mm_sub_ps(iy1,jy2);
 219             dz12             = _mm_sub_ps(iz1,jz2);
 220             dx13             = _mm_sub_ps(ix1,jx3);
 221             dy13             = _mm_sub_ps(iy1,jy3);
 222             dz13             = _mm_sub_ps(iz1,jz3);
 223             dx21             = _mm_sub_ps(ix2,jx1);
 224             dy21             = _mm_sub_ps(iy2,jy1);
 225             dz21             = _mm_sub_ps(iz2,jz1);
 226             dx22             = _mm_sub_ps(ix2,jx2);
 227             dy22             = _mm_sub_ps(iy2,jy2);
 228             dz22             = _mm_sub_ps(iz2,jz2);
 229             dx23             = _mm_sub_ps(ix2,jx3);
 230             dy23             = _mm_sub_ps(iy2,jy3);
 231             dz23             = _mm_sub_ps(iz2,jz3);
 232             dx31             = _mm_sub_ps(ix3,jx1);
 233             dy31             = _mm_sub_ps(iy3,jy1);
 234             dz31             = _mm_sub_ps(iz3,jz1);
 235             dx32             = _mm_sub_ps(ix3,jx2);
 236             dy32             = _mm_sub_ps(iy3,jy2);
 237             dz32             = _mm_sub_ps(iz3,jz2);
 238             dx33             = _mm_sub_ps(ix3,jx3);
 239             dy33             = _mm_sub_ps(iy3,jy3);
 240             dz33             = _mm_sub_ps(iz3,jz3);
 241
 242             /* Calculate squared distance and things based on it */
 243             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 244             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 245             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 246             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 247             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 248             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 249             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 250             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 251             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 252
 253             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 254             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 255             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 256             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 257             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 258             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 259             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 260             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 261             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 262
 263             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 264             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 265             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 266             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 267             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 268             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 269             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 270             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 271             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 272
 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             fjx3             = _mm_setzero_ps();
 280             fjy3             = _mm_setzero_ps();
 281             fjz3             = _mm_setzero_ps();
 282
 283             /**************************
 284              * CALCULATE INTERACTIONS *
 285              **************************/
 286
 287             /* REACTION-FIELD ELECTROSTATICS */
 288             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 289             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 290
 291             /* Update potential sum for this i atom from the interaction with this j atom. */
 292             velecsum         = _mm_add_ps(velecsum,velec);
 293
 294             fscal            = felec;
 295
 296             /* Calculate temporary vectorial force */
 297             tx               = _mm_mul_ps(fscal,dx11);
 298             ty               = _mm_mul_ps(fscal,dy11);
 299             tz               = _mm_mul_ps(fscal,dz11);
 300
 301             /* Update vectorial force */
 302             fix1             = _mm_add_ps(fix1,tx);
 303             fiy1             = _mm_add_ps(fiy1,ty);
 304             fiz1             = _mm_add_ps(fiz1,tz);
 305
 306             fjx1             = _mm_add_ps(fjx1,tx);
 307             fjy1             = _mm_add_ps(fjy1,ty);
 308             fjz1             = _mm_add_ps(fjz1,tz);
 309
 310             /**************************
 311              * CALCULATE INTERACTIONS *
 312              **************************/
 313
 314             /* REACTION-FIELD ELECTROSTATICS */
 315             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 316             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 317
 318             /* Update potential sum for this i atom from the interaction with this j atom. */
 319             velecsum         = _mm_add_ps(velecsum,velec);
 320
 321             fscal            = felec;
 322
 323             /* Calculate temporary vectorial force */
 324             tx               = _mm_mul_ps(fscal,dx12);
 325             ty               = _mm_mul_ps(fscal,dy12);
 326             tz               = _mm_mul_ps(fscal,dz12);
 327
 328             /* Update vectorial force */
 329             fix1             = _mm_add_ps(fix1,tx);
 330             fiy1             = _mm_add_ps(fiy1,ty);
 331             fiz1             = _mm_add_ps(fiz1,tz);
 332
 333             fjx2             = _mm_add_ps(fjx2,tx);
 334             fjy2             = _mm_add_ps(fjy2,ty);
 335             fjz2             = _mm_add_ps(fjz2,tz);
 336
 337             /**************************
 338              * CALCULATE INTERACTIONS *
 339              **************************/
 340
 341             /* REACTION-FIELD ELECTROSTATICS */
 342             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_add_ps(rinv13,_mm_mul_ps(krf,rsq13)),crf));
 343             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
 344
 345             /* Update potential sum for this i atom from the interaction with this j atom. */
 346             velecsum         = _mm_add_ps(velecsum,velec);
 347
 348             fscal            = felec;
 349
 350             /* Calculate temporary vectorial force */
 351             tx               = _mm_mul_ps(fscal,dx13);
 352             ty               = _mm_mul_ps(fscal,dy13);
 353             tz               = _mm_mul_ps(fscal,dz13);
 354
 355             /* Update vectorial force */
 356             fix1             = _mm_add_ps(fix1,tx);
 357             fiy1             = _mm_add_ps(fiy1,ty);
 358             fiz1             = _mm_add_ps(fiz1,tz);
 359
 360             fjx3             = _mm_add_ps(fjx3,tx);
 361             fjy3             = _mm_add_ps(fjy3,ty);
 362             fjz3             = _mm_add_ps(fjz3,tz);
 363
 364             /**************************
 365              * CALCULATE INTERACTIONS *
 366              **************************/
 367
 368             /* REACTION-FIELD ELECTROSTATICS */
 369             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 370             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 371
 372             /* Update potential sum for this i atom from the interaction with this j atom. */
 373             velecsum         = _mm_add_ps(velecsum,velec);
 374
 375             fscal            = felec;
 376
 377             /* Calculate temporary vectorial force */
 378             tx               = _mm_mul_ps(fscal,dx21);
 379             ty               = _mm_mul_ps(fscal,dy21);
 380             tz               = _mm_mul_ps(fscal,dz21);
 381
 382             /* Update vectorial force */
 383             fix2             = _mm_add_ps(fix2,tx);
 384             fiy2             = _mm_add_ps(fiy2,ty);
 385             fiz2             = _mm_add_ps(fiz2,tz);
 386
 387             fjx1             = _mm_add_ps(fjx1,tx);
 388             fjy1             = _mm_add_ps(fjy1,ty);
 389             fjz1             = _mm_add_ps(fjz1,tz);
 390
 391             /**************************
 392              * CALCULATE INTERACTIONS *
 393              **************************/
 394
 395             /* REACTION-FIELD ELECTROSTATICS */
 396             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 397             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 398
 399             /* Update potential sum for this i atom from the interaction with this j atom. */
 400             velecsum         = _mm_add_ps(velecsum,velec);
 401
 402             fscal            = felec;
 403
 404             /* Calculate temporary vectorial force */
 405             tx               = _mm_mul_ps(fscal,dx22);
 406             ty               = _mm_mul_ps(fscal,dy22);
 407             tz               = _mm_mul_ps(fscal,dz22);
 408
 409             /* Update vectorial force */
 410             fix2             = _mm_add_ps(fix2,tx);
 411             fiy2             = _mm_add_ps(fiy2,ty);
 412             fiz2             = _mm_add_ps(fiz2,tz);
 413
 414             fjx2             = _mm_add_ps(fjx2,tx);
 415             fjy2             = _mm_add_ps(fjy2,ty);
 416             fjz2             = _mm_add_ps(fjz2,tz);
 417
 418             /**************************
 419              * CALCULATE INTERACTIONS *
 420              **************************/
 421
 422             /* REACTION-FIELD ELECTROSTATICS */
 423             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_add_ps(rinv23,_mm_mul_ps(krf,rsq23)),crf));
 424             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
 425
 426             /* Update potential sum for this i atom from the interaction with this j atom. */
 427             velecsum         = _mm_add_ps(velecsum,velec);
 428
 429             fscal            = felec;
 430
 431             /* Calculate temporary vectorial force */
 432             tx               = _mm_mul_ps(fscal,dx23);
 433             ty               = _mm_mul_ps(fscal,dy23);
 434             tz               = _mm_mul_ps(fscal,dz23);
 435
 436             /* Update vectorial force */
 437             fix2             = _mm_add_ps(fix2,tx);
 438             fiy2             = _mm_add_ps(fiy2,ty);
 439             fiz2             = _mm_add_ps(fiz2,tz);
 440
 441             fjx3             = _mm_add_ps(fjx3,tx);
 442             fjy3             = _mm_add_ps(fjy3,ty);
 443             fjz3             = _mm_add_ps(fjz3,tz);
 444
 445             /**************************
 446              * CALCULATE INTERACTIONS *
 447              **************************/
 448
 449             /* REACTION-FIELD ELECTROSTATICS */
 450             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_add_ps(rinv31,_mm_mul_ps(krf,rsq31)),crf));
 451             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
 452
 453             /* Update potential sum for this i atom from the interaction with this j atom. */
 454             velecsum         = _mm_add_ps(velecsum,velec);
 455
 456             fscal            = felec;
 457
 458             /* Calculate temporary vectorial force */
 459             tx               = _mm_mul_ps(fscal,dx31);
 460             ty               = _mm_mul_ps(fscal,dy31);
 461             tz               = _mm_mul_ps(fscal,dz31);
 462
 463             /* Update vectorial force */
 464             fix3             = _mm_add_ps(fix3,tx);
 465             fiy3             = _mm_add_ps(fiy3,ty);
 466             fiz3             = _mm_add_ps(fiz3,tz);
 467
 468             fjx1             = _mm_add_ps(fjx1,tx);
 469             fjy1             = _mm_add_ps(fjy1,ty);
 470             fjz1             = _mm_add_ps(fjz1,tz);
 471
 472             /**************************
 473              * CALCULATE INTERACTIONS *
 474              **************************/
 475
 476             /* REACTION-FIELD ELECTROSTATICS */
 477             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_add_ps(rinv32,_mm_mul_ps(krf,rsq32)),crf));
 478             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
 479
 480             /* Update potential sum for this i atom from the interaction with this j atom. */
 481             velecsum         = _mm_add_ps(velecsum,velec);
 482
 483             fscal            = felec;
 484
 485             /* Calculate temporary vectorial force */
 486             tx               = _mm_mul_ps(fscal,dx32);
 487             ty               = _mm_mul_ps(fscal,dy32);
 488             tz               = _mm_mul_ps(fscal,dz32);
 489
 490             /* Update vectorial force */
 491             fix3             = _mm_add_ps(fix3,tx);
 492             fiy3             = _mm_add_ps(fiy3,ty);
 493             fiz3             = _mm_add_ps(fiz3,tz);
 494
 495             fjx2             = _mm_add_ps(fjx2,tx);
 496             fjy2             = _mm_add_ps(fjy2,ty);
 497             fjz2             = _mm_add_ps(fjz2,tz);
 498
 499             /**************************
 500              * CALCULATE INTERACTIONS *
 501              **************************/
 502
 503             /* REACTION-FIELD ELECTROSTATICS */
 504             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_add_ps(rinv33,_mm_mul_ps(krf,rsq33)),crf));
 505             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
 506
 507             /* Update potential sum for this i atom from the interaction with this j atom. */
 508             velecsum         = _mm_add_ps(velecsum,velec);
 509
 510             fscal            = felec;
 511
 512             /* Calculate temporary vectorial force */
 513             tx               = _mm_mul_ps(fscal,dx33);
 514             ty               = _mm_mul_ps(fscal,dy33);
 515             tz               = _mm_mul_ps(fscal,dz33);
 516
 517             /* Update vectorial force */
 518             fix3             = _mm_add_ps(fix3,tx);
 519             fiy3             = _mm_add_ps(fiy3,ty);
 520             fiz3             = _mm_add_ps(fiz3,tz);
 521
 522             fjx3             = _mm_add_ps(fjx3,tx);
 523             fjy3             = _mm_add_ps(fjy3,ty);
 524             fjz3             = _mm_add_ps(fjz3,tz);
 525
 526             fjptrA             = f+j_coord_offsetA;
 527             fjptrB             = f+j_coord_offsetB;
 528             fjptrC             = f+j_coord_offsetC;
 529             fjptrD             = f+j_coord_offsetD;
 530
 531             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
 532                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 533
 534             /* Inner loop uses 288 flops */
 535         }
 536
 537         if(jidx<j_index_end)
 538         {
 539
 540             /* Get j neighbor index, and coordinate index */
 541             jnrlistA         = jjnr[jidx];
 542             jnrlistB         = jjnr[jidx+1];
 543             jnrlistC         = jjnr[jidx+2];
 544             jnrlistD         = jjnr[jidx+3];
 545             /* Sign of each element will be negative for non-real atoms.
 546              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 547              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 548              */
 549             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 550             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 551             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 552             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 553             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 554             j_coord_offsetA  = DIM*jnrA;
 555             j_coord_offsetB  = DIM*jnrB;
 556             j_coord_offsetC  = DIM*jnrC;
 557             j_coord_offsetD  = DIM*jnrD;
 558
 559             /* load j atom coordinates */
 560             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 561                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 562                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 563
 564             /* Calculate displacement vector */
 565             dx11             = _mm_sub_ps(ix1,jx1);
 566             dy11             = _mm_sub_ps(iy1,jy1);
 567             dz11             = _mm_sub_ps(iz1,jz1);
 568             dx12             = _mm_sub_ps(ix1,jx2);
 569             dy12             = _mm_sub_ps(iy1,jy2);
 570             dz12             = _mm_sub_ps(iz1,jz2);
 571             dx13             = _mm_sub_ps(ix1,jx3);
 572             dy13             = _mm_sub_ps(iy1,jy3);
 573             dz13             = _mm_sub_ps(iz1,jz3);
 574             dx21             = _mm_sub_ps(ix2,jx1);
 575             dy21             = _mm_sub_ps(iy2,jy1);
 576             dz21             = _mm_sub_ps(iz2,jz1);
 577             dx22             = _mm_sub_ps(ix2,jx2);
 578             dy22             = _mm_sub_ps(iy2,jy2);
 579             dz22             = _mm_sub_ps(iz2,jz2);
 580             dx23             = _mm_sub_ps(ix2,jx3);
 581             dy23             = _mm_sub_ps(iy2,jy3);
 582             dz23             = _mm_sub_ps(iz2,jz3);
 583             dx31             = _mm_sub_ps(ix3,jx1);
 584             dy31             = _mm_sub_ps(iy3,jy1);
 585             dz31             = _mm_sub_ps(iz3,jz1);
 586             dx32             = _mm_sub_ps(ix3,jx2);
 587             dy32             = _mm_sub_ps(iy3,jy2);
 588             dz32             = _mm_sub_ps(iz3,jz2);
 589             dx33             = _mm_sub_ps(ix3,jx3);
 590             dy33             = _mm_sub_ps(iy3,jy3);
 591             dz33             = _mm_sub_ps(iz3,jz3);
 592
 593             /* Calculate squared distance and things based on it */
 594             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 595             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 596             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 597             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 598             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 599             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 600             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 601             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 602             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 603
 604             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 605             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 606             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 607             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 608             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 609             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 610             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 611             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 612             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 613
 614             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 615             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 616             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 617             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 618             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 619             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 620             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 621             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 622             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 623
 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             fjx3             = _mm_setzero_ps();
 631             fjy3             = _mm_setzero_ps();
 632             fjz3             = _mm_setzero_ps();
 633
 634             /**************************
 635              * CALCULATE INTERACTIONS *
 636              **************************/
 637
 638             /* REACTION-FIELD ELECTROSTATICS */
 639             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 640             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 641
 642             /* Update potential sum for this i atom from the interaction with this j atom. */
 643             velec            = _mm_andnot_ps(dummy_mask,velec);
 644             velecsum         = _mm_add_ps(velecsum,velec);
 645
 646             fscal            = felec;
 647
 648             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 649
 650             /* Calculate temporary vectorial force */
 651             tx               = _mm_mul_ps(fscal,dx11);
 652             ty               = _mm_mul_ps(fscal,dy11);
 653             tz               = _mm_mul_ps(fscal,dz11);
 654
 655             /* Update vectorial force */
 656             fix1             = _mm_add_ps(fix1,tx);
 657             fiy1             = _mm_add_ps(fiy1,ty);
 658             fiz1             = _mm_add_ps(fiz1,tz);
 659
 660             fjx1             = _mm_add_ps(fjx1,tx);
 661             fjy1             = _mm_add_ps(fjy1,ty);
 662             fjz1             = _mm_add_ps(fjz1,tz);
 663
 664             /**************************
 665              * CALCULATE INTERACTIONS *
 666              **************************/
 667
 668             /* REACTION-FIELD ELECTROSTATICS */
 669             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 670             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 671
 672             /* Update potential sum for this i atom from the interaction with this j atom. */
 673             velec            = _mm_andnot_ps(dummy_mask,velec);
 674             velecsum         = _mm_add_ps(velecsum,velec);
 675
 676             fscal            = felec;
 677
 678             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 679
 680             /* Calculate temporary vectorial force */
 681             tx               = _mm_mul_ps(fscal,dx12);
 682             ty               = _mm_mul_ps(fscal,dy12);
 683             tz               = _mm_mul_ps(fscal,dz12);
 684
 685             /* Update vectorial force */
 686             fix1             = _mm_add_ps(fix1,tx);
 687             fiy1             = _mm_add_ps(fiy1,ty);
 688             fiz1             = _mm_add_ps(fiz1,tz);
 689
 690             fjx2             = _mm_add_ps(fjx2,tx);
 691             fjy2             = _mm_add_ps(fjy2,ty);
 692             fjz2             = _mm_add_ps(fjz2,tz);
 693
 694             /**************************
 695              * CALCULATE INTERACTIONS *
 696              **************************/
 697
 698             /* REACTION-FIELD ELECTROSTATICS */
 699             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_add_ps(rinv13,_mm_mul_ps(krf,rsq13)),crf));
 700             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
 701
 702             /* Update potential sum for this i atom from the interaction with this j atom. */
 703             velec            = _mm_andnot_ps(dummy_mask,velec);
 704             velecsum         = _mm_add_ps(velecsum,velec);
 705
 706             fscal            = felec;
 707
 708             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 709
 710             /* Calculate temporary vectorial force */
 711             tx               = _mm_mul_ps(fscal,dx13);
 712             ty               = _mm_mul_ps(fscal,dy13);
 713             tz               = _mm_mul_ps(fscal,dz13);
 714
 715             /* Update vectorial force */
 716             fix1             = _mm_add_ps(fix1,tx);
 717             fiy1             = _mm_add_ps(fiy1,ty);
 718             fiz1             = _mm_add_ps(fiz1,tz);
 719
 720             fjx3             = _mm_add_ps(fjx3,tx);
 721             fjy3             = _mm_add_ps(fjy3,ty);
 722             fjz3             = _mm_add_ps(fjz3,tz);
 723
 724             /**************************
 725              * CALCULATE INTERACTIONS *
 726              **************************/
 727
 728             /* REACTION-FIELD ELECTROSTATICS */
 729             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 730             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 731
 732             /* Update potential sum for this i atom from the interaction with this j atom. */
 733             velec            = _mm_andnot_ps(dummy_mask,velec);
 734             velecsum         = _mm_add_ps(velecsum,velec);
 735
 736             fscal            = felec;
 737
 738             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 739
 740             /* Calculate temporary vectorial force */
 741             tx               = _mm_mul_ps(fscal,dx21);
 742             ty               = _mm_mul_ps(fscal,dy21);
 743             tz               = _mm_mul_ps(fscal,dz21);
 744
 745             /* Update vectorial force */
 746             fix2             = _mm_add_ps(fix2,tx);
 747             fiy2             = _mm_add_ps(fiy2,ty);
 748             fiz2             = _mm_add_ps(fiz2,tz);
 749
 750             fjx1             = _mm_add_ps(fjx1,tx);
 751             fjy1             = _mm_add_ps(fjy1,ty);
 752             fjz1             = _mm_add_ps(fjz1,tz);
 753
 754             /**************************
 755              * CALCULATE INTERACTIONS *
 756              **************************/
 757
 758             /* REACTION-FIELD ELECTROSTATICS */
 759             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 760             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 761
 762             /* Update potential sum for this i atom from the interaction with this j atom. */
 763             velec            = _mm_andnot_ps(dummy_mask,velec);
 764             velecsum         = _mm_add_ps(velecsum,velec);
 765
 766             fscal            = felec;
 767
 768             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 769
 770             /* Calculate temporary vectorial force */
 771             tx               = _mm_mul_ps(fscal,dx22);
 772             ty               = _mm_mul_ps(fscal,dy22);
 773             tz               = _mm_mul_ps(fscal,dz22);
 774
 775             /* Update vectorial force */
 776             fix2             = _mm_add_ps(fix2,tx);
 777             fiy2             = _mm_add_ps(fiy2,ty);
 778             fiz2             = _mm_add_ps(fiz2,tz);
 779
 780             fjx2             = _mm_add_ps(fjx2,tx);
 781             fjy2             = _mm_add_ps(fjy2,ty);
 782             fjz2             = _mm_add_ps(fjz2,tz);
 783
 784             /**************************
 785              * CALCULATE INTERACTIONS *
 786              **************************/
 787
 788             /* REACTION-FIELD ELECTROSTATICS */
 789             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_add_ps(rinv23,_mm_mul_ps(krf,rsq23)),crf));
 790             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
 791
 792             /* Update potential sum for this i atom from the interaction with this j atom. */
 793             velec            = _mm_andnot_ps(dummy_mask,velec);
 794             velecsum         = _mm_add_ps(velecsum,velec);
 795
 796             fscal            = felec;
 797
 798             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 799
 800             /* Calculate temporary vectorial force */
 801             tx               = _mm_mul_ps(fscal,dx23);
 802             ty               = _mm_mul_ps(fscal,dy23);
 803             tz               = _mm_mul_ps(fscal,dz23);
 804
 805             /* Update vectorial force */
 806             fix2             = _mm_add_ps(fix2,tx);
 807             fiy2             = _mm_add_ps(fiy2,ty);
 808             fiz2             = _mm_add_ps(fiz2,tz);
 809
 810             fjx3             = _mm_add_ps(fjx3,tx);
 811             fjy3             = _mm_add_ps(fjy3,ty);
 812             fjz3             = _mm_add_ps(fjz3,tz);
 813
 814             /**************************
 815              * CALCULATE INTERACTIONS *
 816              **************************/
 817
 818             /* REACTION-FIELD ELECTROSTATICS */
 819             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_add_ps(rinv31,_mm_mul_ps(krf,rsq31)),crf));
 820             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
 821
 822             /* Update potential sum for this i atom from the interaction with this j atom. */
 823             velec            = _mm_andnot_ps(dummy_mask,velec);
 824             velecsum         = _mm_add_ps(velecsum,velec);
 825
 826             fscal            = felec;
 827
 828             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 829
 830             /* Calculate temporary vectorial force */
 831             tx               = _mm_mul_ps(fscal,dx31);
 832             ty               = _mm_mul_ps(fscal,dy31);
 833             tz               = _mm_mul_ps(fscal,dz31);
 834
 835             /* Update vectorial force */
 836             fix3             = _mm_add_ps(fix3,tx);
 837             fiy3             = _mm_add_ps(fiy3,ty);
 838             fiz3             = _mm_add_ps(fiz3,tz);
 839
 840             fjx1             = _mm_add_ps(fjx1,tx);
 841             fjy1             = _mm_add_ps(fjy1,ty);
 842             fjz1             = _mm_add_ps(fjz1,tz);
 843
 844             /**************************
 845              * CALCULATE INTERACTIONS *
 846              **************************/
 847
 848             /* REACTION-FIELD ELECTROSTATICS */
 849             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_add_ps(rinv32,_mm_mul_ps(krf,rsq32)),crf));
 850             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
 851
 852             /* Update potential sum for this i atom from the interaction with this j atom. */
 853             velec            = _mm_andnot_ps(dummy_mask,velec);
 854             velecsum         = _mm_add_ps(velecsum,velec);
 855
 856             fscal            = felec;
 857
 858             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 859
 860             /* Calculate temporary vectorial force */
 861             tx               = _mm_mul_ps(fscal,dx32);
 862             ty               = _mm_mul_ps(fscal,dy32);
 863             tz               = _mm_mul_ps(fscal,dz32);
 864
 865             /* Update vectorial force */
 866             fix3             = _mm_add_ps(fix3,tx);
 867             fiy3             = _mm_add_ps(fiy3,ty);
 868             fiz3             = _mm_add_ps(fiz3,tz);
 869
 870             fjx2             = _mm_add_ps(fjx2,tx);
 871             fjy2             = _mm_add_ps(fjy2,ty);
 872             fjz2             = _mm_add_ps(fjz2,tz);
 873
 874             /**************************
 875              * CALCULATE INTERACTIONS *
 876              **************************/
 877
 878             /* REACTION-FIELD ELECTROSTATICS */
 879             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_add_ps(rinv33,_mm_mul_ps(krf,rsq33)),crf));
 880             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
 881
 882             /* Update potential sum for this i atom from the interaction with this j atom. */
 883             velec            = _mm_andnot_ps(dummy_mask,velec);
 884             velecsum         = _mm_add_ps(velecsum,velec);
 885
 886             fscal            = felec;
 887
 888             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 889
 890             /* Calculate temporary vectorial force */
 891             tx               = _mm_mul_ps(fscal,dx33);
 892             ty               = _mm_mul_ps(fscal,dy33);
 893             tz               = _mm_mul_ps(fscal,dz33);
 894
 895             /* Update vectorial force */
 896             fix3             = _mm_add_ps(fix3,tx);
 897             fiy3             = _mm_add_ps(fiy3,ty);
 898             fiz3             = _mm_add_ps(fiz3,tz);
 899
 900             fjx3             = _mm_add_ps(fjx3,tx);
 901             fjy3             = _mm_add_ps(fjy3,ty);
 902             fjz3             = _mm_add_ps(fjz3,tz);
 903
 904             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 905             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 906             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 907             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 908
 909             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
 910                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 911
 912             /* Inner loop uses 288 flops */
 913         }
 914
 915         /* End of innermost loop */
 916
 917         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 918                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
 919
 920         ggid                        = gid[iidx];
 921         /* Update potential energies */
 922         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 923
 924         /* Increment number of inner iterations */
 925         inneriter                  += j_index_end - j_index_start;
 926
 927         /* Outer loop uses 19 flops */
 928     }
 929
 930     /* Increment number of outer iterations */
 931     outeriter        += nri;
 932
 933     /* Update outer/inner flops */
 934
 935     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*288);
 936 }
 937 /*
 938  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwNone_GeomW4W4_F_sse2_single
 939  * Electrostatics interaction: ReactionField
 940  * VdW interaction:            None
 941  * Geometry:                   Water4-Water4
 942  * Calculate force/pot:        Force
 943  */
 944 void
 945 nb_kernel_ElecRF_VdwNone_GeomW4W4_F_sse2_single
 946                     (t_nblist                    * gmx_restrict       nlist,
 947                      rvec                        * gmx_restrict          xx,
 948                      rvec                        * gmx_restrict          ff,
 949                      t_forcerec                  * gmx_restrict          fr,
 950                      t_mdatoms                   * gmx_restrict     mdatoms,
 951                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 952                      t_nrnb                      * gmx_restrict        nrnb)
 953 {
 954     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 955      * just 0 for non-waters.
 956      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 957      * jnr indices corresponding to data put in the four positions in the SIMD register.
 958      */
 959     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 960     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 961     int              jnrA,jnrB,jnrC,jnrD;
 962     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 963     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 964     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 965     real             rcutoff_scalar;
 966     real             *shiftvec,*fshift,*x,*f;
 967     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 968     real             scratch[4*DIM];
 969     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 970     int              vdwioffset1;
 971     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 972     int              vdwioffset2;
 973     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 974     int              vdwioffset3;
 975     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 976     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
 977     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 978     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
 979     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 980     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
 981     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 982     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 983     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 984     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 985     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 986     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 987     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 988     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 989     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 990     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 991     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 992     real             *charge;
 993     __m128           dummy_mask,cutoff_mask;
 994     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 995     __m128           one     = _mm_set1_ps(1.0);
 996     __m128           two     = _mm_set1_ps(2.0);
 997     x                = xx[0];
 998     f                = ff[0];
 999
1000     nri              = nlist->nri;
1001     iinr             = nlist->iinr;
1002     jindex           = nlist->jindex;
1003     jjnr             = nlist->jjnr;
1004     shiftidx         = nlist->shift;
1005     gid              = nlist->gid;
1006     shiftvec         = fr->shift_vec[0];
1007     fshift           = fr->fshift[0];
1008     facel            = _mm_set1_ps(fr->epsfac);
1009     charge           = mdatoms->chargeA;
1010     krf              = _mm_set1_ps(fr->ic->k_rf);
1011     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1012     crf              = _mm_set1_ps(fr->ic->c_rf);
1013
1014     /* Setup water-specific parameters */
1015     inr              = nlist->iinr[0];
1016     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1017     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1018     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1019
1020     jq1              = _mm_set1_ps(charge[inr+1]);
1021     jq2              = _mm_set1_ps(charge[inr+2]);
1022     jq3              = _mm_set1_ps(charge[inr+3]);
1023     qq11             = _mm_mul_ps(iq1,jq1);
1024     qq12             = _mm_mul_ps(iq1,jq2);
1025     qq13             = _mm_mul_ps(iq1,jq3);
1026     qq21             = _mm_mul_ps(iq2,jq1);
1027     qq22             = _mm_mul_ps(iq2,jq2);
1028     qq23             = _mm_mul_ps(iq2,jq3);
1029     qq31             = _mm_mul_ps(iq3,jq1);
1030     qq32             = _mm_mul_ps(iq3,jq2);
1031     qq33             = _mm_mul_ps(iq3,jq3);
1032
1033     /* Avoid stupid compiler warnings */
1034     jnrA = jnrB = jnrC = jnrD = 0;
1035     j_coord_offsetA = 0;
1036     j_coord_offsetB = 0;
1037     j_coord_offsetC = 0;
1038     j_coord_offsetD = 0;
1039
1040     outeriter        = 0;
1041     inneriter        = 0;
1042
1043     for(iidx=0;iidx<4*DIM;iidx++)
1044     {
1045         scratch[iidx] = 0.0;
1046     }
1047
1048     /* Start outer loop over neighborlists */
1049     for(iidx=0; iidx<nri; iidx++)
1050     {
1051         /* Load shift vector for this list */
1052         i_shift_offset   = DIM*shiftidx[iidx];
1053
1054         /* Load limits for loop over neighbors */
1055         j_index_start    = jindex[iidx];
1056         j_index_end      = jindex[iidx+1];
1057
1058         /* Get outer coordinate index */
1059         inr              = iinr[iidx];
1060         i_coord_offset   = DIM*inr;
1061
1062         /* Load i particle coords and add shift vector */
1063         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
1064                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1065
1066         fix1             = _mm_setzero_ps();
1067         fiy1             = _mm_setzero_ps();
1068         fiz1             = _mm_setzero_ps();
1069         fix2             = _mm_setzero_ps();
1070         fiy2             = _mm_setzero_ps();
1071         fiz2             = _mm_setzero_ps();
1072         fix3             = _mm_setzero_ps();
1073         fiy3             = _mm_setzero_ps();
1074         fiz3             = _mm_setzero_ps();
1075
1076         /* Start inner kernel loop */
1077         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1078         {
1079
1080             /* Get j neighbor index, and coordinate index */
1081             jnrA             = jjnr[jidx];
1082             jnrB             = jjnr[jidx+1];
1083             jnrC             = jjnr[jidx+2];
1084             jnrD             = jjnr[jidx+3];
1085             j_coord_offsetA  = DIM*jnrA;
1086             j_coord_offsetB  = DIM*jnrB;
1087             j_coord_offsetC  = DIM*jnrC;
1088             j_coord_offsetD  = DIM*jnrD;
1089
1090             /* load j atom coordinates */
1091             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1092                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1093                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1094
1095             /* Calculate displacement vector */
1096             dx11             = _mm_sub_ps(ix1,jx1);
1097             dy11             = _mm_sub_ps(iy1,jy1);
1098             dz11             = _mm_sub_ps(iz1,jz1);
1099             dx12             = _mm_sub_ps(ix1,jx2);
1100             dy12             = _mm_sub_ps(iy1,jy2);
1101             dz12             = _mm_sub_ps(iz1,jz2);
1102             dx13             = _mm_sub_ps(ix1,jx3);
1103             dy13             = _mm_sub_ps(iy1,jy3);
1104             dz13             = _mm_sub_ps(iz1,jz3);
1105             dx21             = _mm_sub_ps(ix2,jx1);
1106             dy21             = _mm_sub_ps(iy2,jy1);
1107             dz21             = _mm_sub_ps(iz2,jz1);
1108             dx22             = _mm_sub_ps(ix2,jx2);
1109             dy22             = _mm_sub_ps(iy2,jy2);
1110             dz22             = _mm_sub_ps(iz2,jz2);
1111             dx23             = _mm_sub_ps(ix2,jx3);
1112             dy23             = _mm_sub_ps(iy2,jy3);
1113             dz23             = _mm_sub_ps(iz2,jz3);
1114             dx31             = _mm_sub_ps(ix3,jx1);
1115             dy31             = _mm_sub_ps(iy3,jy1);
1116             dz31             = _mm_sub_ps(iz3,jz1);
1117             dx32             = _mm_sub_ps(ix3,jx2);
1118             dy32             = _mm_sub_ps(iy3,jy2);
1119             dz32             = _mm_sub_ps(iz3,jz2);
1120             dx33             = _mm_sub_ps(ix3,jx3);
1121             dy33             = _mm_sub_ps(iy3,jy3);
1122             dz33             = _mm_sub_ps(iz3,jz3);
1123
1124             /* Calculate squared distance and things based on it */
1125             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1126             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1127             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1128             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1129             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1130             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1131             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1132             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1133             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1134
1135             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1136             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1137             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1138             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1139             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1140             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1141             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1142             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1143             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1144
1145             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1146             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1147             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1148             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1149             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1150             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1151             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1152             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1153             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1154
1155             fjx1             = _mm_setzero_ps();
1156             fjy1             = _mm_setzero_ps();
1157             fjz1             = _mm_setzero_ps();
1158             fjx2             = _mm_setzero_ps();
1159             fjy2             = _mm_setzero_ps();
1160             fjz2             = _mm_setzero_ps();
1161             fjx3             = _mm_setzero_ps();
1162             fjy3             = _mm_setzero_ps();
1163             fjz3             = _mm_setzero_ps();
1164
1165             /**************************
1166              * CALCULATE INTERACTIONS *
1167              **************************/
1168
1169             /* REACTION-FIELD ELECTROSTATICS */
1170             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1171
1172             fscal            = felec;
1173
1174             /* Calculate temporary vectorial force */
1175             tx               = _mm_mul_ps(fscal,dx11);
1176             ty               = _mm_mul_ps(fscal,dy11);
1177             tz               = _mm_mul_ps(fscal,dz11);
1178
1179             /* Update vectorial force */
1180             fix1             = _mm_add_ps(fix1,tx);
1181             fiy1             = _mm_add_ps(fiy1,ty);
1182             fiz1             = _mm_add_ps(fiz1,tz);
1183
1184             fjx1             = _mm_add_ps(fjx1,tx);
1185             fjy1             = _mm_add_ps(fjy1,ty);
1186             fjz1             = _mm_add_ps(fjz1,tz);
1187
1188             /**************************
1189              * CALCULATE INTERACTIONS *
1190              **************************/
1191
1192             /* REACTION-FIELD ELECTROSTATICS */
1193             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1194
1195             fscal            = felec;
1196
1197             /* Calculate temporary vectorial force */
1198             tx               = _mm_mul_ps(fscal,dx12);
1199             ty               = _mm_mul_ps(fscal,dy12);
1200             tz               = _mm_mul_ps(fscal,dz12);
1201
1202             /* Update vectorial force */
1203             fix1             = _mm_add_ps(fix1,tx);
1204             fiy1             = _mm_add_ps(fiy1,ty);
1205             fiz1             = _mm_add_ps(fiz1,tz);
1206
1207             fjx2             = _mm_add_ps(fjx2,tx);
1208             fjy2             = _mm_add_ps(fjy2,ty);
1209             fjz2             = _mm_add_ps(fjz2,tz);
1210
1211             /**************************
1212              * CALCULATE INTERACTIONS *
1213              **************************/
1214
1215             /* REACTION-FIELD ELECTROSTATICS */
1216             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
1217
1218             fscal            = felec;
1219
1220             /* Calculate temporary vectorial force */
1221             tx               = _mm_mul_ps(fscal,dx13);
1222             ty               = _mm_mul_ps(fscal,dy13);
1223             tz               = _mm_mul_ps(fscal,dz13);
1224
1225             /* Update vectorial force */
1226             fix1             = _mm_add_ps(fix1,tx);
1227             fiy1             = _mm_add_ps(fiy1,ty);
1228             fiz1             = _mm_add_ps(fiz1,tz);
1229
1230             fjx3             = _mm_add_ps(fjx3,tx);
1231             fjy3             = _mm_add_ps(fjy3,ty);
1232             fjz3             = _mm_add_ps(fjz3,tz);
1233
1234             /**************************
1235              * CALCULATE INTERACTIONS *
1236              **************************/
1237
1238             /* REACTION-FIELD ELECTROSTATICS */
1239             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1240
1241             fscal            = felec;
1242
1243             /* Calculate temporary vectorial force */
1244             tx               = _mm_mul_ps(fscal,dx21);
1245             ty               = _mm_mul_ps(fscal,dy21);
1246             tz               = _mm_mul_ps(fscal,dz21);
1247
1248             /* Update vectorial force */
1249             fix2             = _mm_add_ps(fix2,tx);
1250             fiy2             = _mm_add_ps(fiy2,ty);
1251             fiz2             = _mm_add_ps(fiz2,tz);
1252
1253             fjx1             = _mm_add_ps(fjx1,tx);
1254             fjy1             = _mm_add_ps(fjy1,ty);
1255             fjz1             = _mm_add_ps(fjz1,tz);
1256
1257             /**************************
1258              * CALCULATE INTERACTIONS *
1259              **************************/
1260
1261             /* REACTION-FIELD ELECTROSTATICS */
1262             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1263
1264             fscal            = felec;
1265
1266             /* Calculate temporary vectorial force */
1267             tx               = _mm_mul_ps(fscal,dx22);
1268             ty               = _mm_mul_ps(fscal,dy22);
1269             tz               = _mm_mul_ps(fscal,dz22);
1270
1271             /* Update vectorial force */
1272             fix2             = _mm_add_ps(fix2,tx);
1273             fiy2             = _mm_add_ps(fiy2,ty);
1274             fiz2             = _mm_add_ps(fiz2,tz);
1275
1276             fjx2             = _mm_add_ps(fjx2,tx);
1277             fjy2             = _mm_add_ps(fjy2,ty);
1278             fjz2             = _mm_add_ps(fjz2,tz);
1279
1280             /**************************
1281              * CALCULATE INTERACTIONS *
1282              **************************/
1283
1284             /* REACTION-FIELD ELECTROSTATICS */
1285             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
1286
1287             fscal            = felec;
1288
1289             /* Calculate temporary vectorial force */
1290             tx               = _mm_mul_ps(fscal,dx23);
1291             ty               = _mm_mul_ps(fscal,dy23);
1292             tz               = _mm_mul_ps(fscal,dz23);
1293
1294             /* Update vectorial force */
1295             fix2             = _mm_add_ps(fix2,tx);
1296             fiy2             = _mm_add_ps(fiy2,ty);
1297             fiz2             = _mm_add_ps(fiz2,tz);
1298
1299             fjx3             = _mm_add_ps(fjx3,tx);
1300             fjy3             = _mm_add_ps(fjy3,ty);
1301             fjz3             = _mm_add_ps(fjz3,tz);
1302
1303             /**************************
1304              * CALCULATE INTERACTIONS *
1305              **************************/
1306
1307             /* REACTION-FIELD ELECTROSTATICS */
1308             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
1309
1310             fscal            = felec;
1311
1312             /* Calculate temporary vectorial force */
1313             tx               = _mm_mul_ps(fscal,dx31);
1314             ty               = _mm_mul_ps(fscal,dy31);
1315             tz               = _mm_mul_ps(fscal,dz31);
1316
1317             /* Update vectorial force */
1318             fix3             = _mm_add_ps(fix3,tx);
1319             fiy3             = _mm_add_ps(fiy3,ty);
1320             fiz3             = _mm_add_ps(fiz3,tz);
1321
1322             fjx1             = _mm_add_ps(fjx1,tx);
1323             fjy1             = _mm_add_ps(fjy1,ty);
1324             fjz1             = _mm_add_ps(fjz1,tz);
1325
1326             /**************************
1327              * CALCULATE INTERACTIONS *
1328              **************************/
1329
1330             /* REACTION-FIELD ELECTROSTATICS */
1331             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1332
1333             fscal            = felec;
1334
1335             /* Calculate temporary vectorial force */
1336             tx               = _mm_mul_ps(fscal,dx32);
1337             ty               = _mm_mul_ps(fscal,dy32);
1338             tz               = _mm_mul_ps(fscal,dz32);
1339
1340             /* Update vectorial force */
1341             fix3             = _mm_add_ps(fix3,tx);
1342             fiy3             = _mm_add_ps(fiy3,ty);
1343             fiz3             = _mm_add_ps(fiz3,tz);
1344
1345             fjx2             = _mm_add_ps(fjx2,tx);
1346             fjy2             = _mm_add_ps(fjy2,ty);
1347             fjz2             = _mm_add_ps(fjz2,tz);
1348
1349             /**************************
1350              * CALCULATE INTERACTIONS *
1351              **************************/
1352
1353             /* REACTION-FIELD ELECTROSTATICS */
1354             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
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           = gmx_mm_invsqrt_ps(rsq11);
1451             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1452             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1453             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1454             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1455             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1456             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1457             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1458             rinv33           = gmx_mm_invsqrt_ps(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             /* REACTION-FIELD ELECTROSTATICS */
1485             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1486
1487             fscal            = felec;
1488
1489             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1490
1491             /* Calculate temporary vectorial force */
1492             tx               = _mm_mul_ps(fscal,dx11);
1493             ty               = _mm_mul_ps(fscal,dy11);
1494             tz               = _mm_mul_ps(fscal,dz11);
1495
1496             /* Update vectorial force */
1497             fix1             = _mm_add_ps(fix1,tx);
1498             fiy1             = _mm_add_ps(fiy1,ty);
1499             fiz1             = _mm_add_ps(fiz1,tz);
1500
1501             fjx1             = _mm_add_ps(fjx1,tx);
1502             fjy1             = _mm_add_ps(fjy1,ty);
1503             fjz1             = _mm_add_ps(fjz1,tz);
1504
1505             /**************************
1506              * CALCULATE INTERACTIONS *
1507              **************************/
1508
1509             /* REACTION-FIELD ELECTROSTATICS */
1510             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1511
1512             fscal            = felec;
1513
1514             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1515
1516             /* Calculate temporary vectorial force */
1517             tx               = _mm_mul_ps(fscal,dx12);
1518             ty               = _mm_mul_ps(fscal,dy12);
1519             tz               = _mm_mul_ps(fscal,dz12);
1520
1521             /* Update vectorial force */
1522             fix1             = _mm_add_ps(fix1,tx);
1523             fiy1             = _mm_add_ps(fiy1,ty);
1524             fiz1             = _mm_add_ps(fiz1,tz);
1525
1526             fjx2             = _mm_add_ps(fjx2,tx);
1527             fjy2             = _mm_add_ps(fjy2,ty);
1528             fjz2             = _mm_add_ps(fjz2,tz);
1529
1530             /**************************
1531              * CALCULATE INTERACTIONS *
1532              **************************/
1533
1534             /* REACTION-FIELD ELECTROSTATICS */
1535             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
1536
1537             fscal            = felec;
1538
1539             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1540
1541             /* Calculate temporary vectorial force */
1542             tx               = _mm_mul_ps(fscal,dx13);
1543             ty               = _mm_mul_ps(fscal,dy13);
1544             tz               = _mm_mul_ps(fscal,dz13);
1545
1546             /* Update vectorial force */
1547             fix1             = _mm_add_ps(fix1,tx);
1548             fiy1             = _mm_add_ps(fiy1,ty);
1549             fiz1             = _mm_add_ps(fiz1,tz);
1550
1551             fjx3             = _mm_add_ps(fjx3,tx);
1552             fjy3             = _mm_add_ps(fjy3,ty);
1553             fjz3             = _mm_add_ps(fjz3,tz);
1554
1555             /**************************
1556              * CALCULATE INTERACTIONS *
1557              **************************/
1558
1559             /* REACTION-FIELD ELECTROSTATICS */
1560             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1561
1562             fscal            = felec;
1563
1564             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1565
1566             /* Calculate temporary vectorial force */
1567             tx               = _mm_mul_ps(fscal,dx21);
1568             ty               = _mm_mul_ps(fscal,dy21);
1569             tz               = _mm_mul_ps(fscal,dz21);
1570
1571             /* Update vectorial force */
1572             fix2             = _mm_add_ps(fix2,tx);
1573             fiy2             = _mm_add_ps(fiy2,ty);
1574             fiz2             = _mm_add_ps(fiz2,tz);
1575
1576             fjx1             = _mm_add_ps(fjx1,tx);
1577             fjy1             = _mm_add_ps(fjy1,ty);
1578             fjz1             = _mm_add_ps(fjz1,tz);
1579
1580             /**************************
1581              * CALCULATE INTERACTIONS *
1582              **************************/
1583
1584             /* REACTION-FIELD ELECTROSTATICS */
1585             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1586
1587             fscal            = felec;
1588
1589             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1590
1591             /* Calculate temporary vectorial force */
1592             tx               = _mm_mul_ps(fscal,dx22);
1593             ty               = _mm_mul_ps(fscal,dy22);
1594             tz               = _mm_mul_ps(fscal,dz22);
1595
1596             /* Update vectorial force */
1597             fix2             = _mm_add_ps(fix2,tx);
1598             fiy2             = _mm_add_ps(fiy2,ty);
1599             fiz2             = _mm_add_ps(fiz2,tz);
1600
1601             fjx2             = _mm_add_ps(fjx2,tx);
1602             fjy2             = _mm_add_ps(fjy2,ty);
1603             fjz2             = _mm_add_ps(fjz2,tz);
1604
1605             /**************************
1606              * CALCULATE INTERACTIONS *
1607              **************************/
1608
1609             /* REACTION-FIELD ELECTROSTATICS */
1610             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
1611
1612             fscal            = felec;
1613
1614             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1615
1616             /* Calculate temporary vectorial force */
1617             tx               = _mm_mul_ps(fscal,dx23);
1618             ty               = _mm_mul_ps(fscal,dy23);
1619             tz               = _mm_mul_ps(fscal,dz23);
1620
1621             /* Update vectorial force */
1622             fix2             = _mm_add_ps(fix2,tx);
1623             fiy2             = _mm_add_ps(fiy2,ty);
1624             fiz2             = _mm_add_ps(fiz2,tz);
1625
1626             fjx3             = _mm_add_ps(fjx3,tx);
1627             fjy3             = _mm_add_ps(fjy3,ty);
1628             fjz3             = _mm_add_ps(fjz3,tz);
1629
1630             /**************************
1631              * CALCULATE INTERACTIONS *
1632              **************************/
1633
1634             /* REACTION-FIELD ELECTROSTATICS */
1635             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
1636
1637             fscal            = felec;
1638
1639             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1640
1641             /* Calculate temporary vectorial force */
1642             tx               = _mm_mul_ps(fscal,dx31);
1643             ty               = _mm_mul_ps(fscal,dy31);
1644             tz               = _mm_mul_ps(fscal,dz31);
1645
1646             /* Update vectorial force */
1647             fix3             = _mm_add_ps(fix3,tx);
1648             fiy3             = _mm_add_ps(fiy3,ty);
1649             fiz3             = _mm_add_ps(fiz3,tz);
1650
1651             fjx1             = _mm_add_ps(fjx1,tx);
1652             fjy1             = _mm_add_ps(fjy1,ty);
1653             fjz1             = _mm_add_ps(fjz1,tz);
1654
1655             /**************************
1656              * CALCULATE INTERACTIONS *
1657              **************************/
1658
1659             /* REACTION-FIELD ELECTROSTATICS */
1660             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1661
1662             fscal            = felec;
1663
1664             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1665
1666             /* Calculate temporary vectorial force */
1667             tx               = _mm_mul_ps(fscal,dx32);
1668             ty               = _mm_mul_ps(fscal,dy32);
1669             tz               = _mm_mul_ps(fscal,dz32);
1670
1671             /* Update vectorial force */
1672             fix3             = _mm_add_ps(fix3,tx);
1673             fiy3             = _mm_add_ps(fiy3,ty);
1674             fiz3             = _mm_add_ps(fiz3,tz);
1675
1676             fjx2             = _mm_add_ps(fjx2,tx);
1677             fjy2             = _mm_add_ps(fjy2,ty);
1678             fjz2             = _mm_add_ps(fjz2,tz);
1679
1680             /**************************
1681              * CALCULATE INTERACTIONS *
1682              **************************/
1683
1684             /* REACTION-FIELD ELECTROSTATICS */
1685             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
1686
1687             fscal            = felec;
1688
1689             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1690
1691             /* Calculate temporary vectorial force */
1692             tx               = _mm_mul_ps(fscal,dx33);
1693             ty               = _mm_mul_ps(fscal,dy33);
1694             tz               = _mm_mul_ps(fscal,dz33);
1695
1696             /* Update vectorial force */
1697             fix3             = _mm_add_ps(fix3,tx);
1698             fiy3             = _mm_add_ps(fiy3,ty);
1699             fiz3             = _mm_add_ps(fiz3,tz);
1700
1701             fjx3             = _mm_add_ps(fjx3,tx);
1702             fjy3             = _mm_add_ps(fjy3,ty);
1703             fjz3             = _mm_add_ps(fjz3,tz);
1704
1705             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1706             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1707             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1708             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1709
1710             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1711                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1712
1713             /* Inner loop uses 243 flops */
1714         }
1715
1716         /* End of innermost loop */
1717
1718         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1719                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1720
1721         /* Increment number of inner iterations */
1722         inneriter                  += j_index_end - j_index_start;
1723
1724         /* Outer loop uses 18 flops */
1725     }
1726
1727     /* Increment number of outer iterations */
1728     outeriter        += nri;
1729
1730     /* Update outer/inner flops */
1731
1732     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*243);
1733 }