src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_ElecRFCut_VdwNone_GeomW4W4_sse4_1_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS sse4_1_single kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_sse4_1_single.h"
  48 #include "kernelutil_x86_sse4_1_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwNone_GeomW4W4_VF_sse4_1_single
  52  * Electrostatics interaction: ReactionField
  53  * VdW interaction:            None
  54  * Geometry:                   Water4-Water4
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecRFCut_VdwNone_GeomW4W4_VF_sse4_1_single
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  83     int              vdwioffset1;
  84     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  85     int              vdwioffset2;
  86     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  87     int              vdwioffset3;
  88     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  89     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  90     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  91     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  92     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  93     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  94     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  95     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  96     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  97     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  98     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  99     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 100     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 101     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 102     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 103     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 104     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 105     real             *charge;
 106     __m128           dummy_mask,cutoff_mask;
 107     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 108     __m128           one     = _mm_set1_ps(1.0);
 109     __m128           two     = _mm_set1_ps(2.0);
 110     x                = xx[0];
 111     f                = ff[0];
 112
 113     nri              = nlist->nri;
 114     iinr             = nlist->iinr;
 115     jindex           = nlist->jindex;
 116     jjnr             = nlist->jjnr;
 117     shiftidx         = nlist->shift;
 118     gid              = nlist->gid;
 119     shiftvec         = fr->shift_vec[0];
 120     fshift           = fr->fshift[0];
 121     facel            = _mm_set1_ps(fr->epsfac);
 122     charge           = mdatoms->chargeA;
 123     krf              = _mm_set1_ps(fr->ic->k_rf);
 124     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 125     crf              = _mm_set1_ps(fr->ic->c_rf);
 126
 127     /* Setup water-specific parameters */
 128     inr              = nlist->iinr[0];
 129     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 130     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 131     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 132
 133     jq1              = _mm_set1_ps(charge[inr+1]);
 134     jq2              = _mm_set1_ps(charge[inr+2]);
 135     jq3              = _mm_set1_ps(charge[inr+3]);
 136     qq11             = _mm_mul_ps(iq1,jq1);
 137     qq12             = _mm_mul_ps(iq1,jq2);
 138     qq13             = _mm_mul_ps(iq1,jq3);
 139     qq21             = _mm_mul_ps(iq2,jq1);
 140     qq22             = _mm_mul_ps(iq2,jq2);
 141     qq23             = _mm_mul_ps(iq2,jq3);
 142     qq31             = _mm_mul_ps(iq3,jq1);
 143     qq32             = _mm_mul_ps(iq3,jq2);
 144     qq33             = _mm_mul_ps(iq3,jq3);
 145
 146     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 147     rcutoff_scalar   = fr->rcoulomb;
 148     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 149     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 150
 151     /* Avoid stupid compiler warnings */
 152     jnrA = jnrB = jnrC = jnrD = 0;
 153     j_coord_offsetA = 0;
 154     j_coord_offsetB = 0;
 155     j_coord_offsetC = 0;
 156     j_coord_offsetD = 0;
 157
 158     outeriter        = 0;
 159     inneriter        = 0;
 160
 161     for(iidx=0;iidx<4*DIM;iidx++)
 162     {
 163         scratch[iidx] = 0.0;
 164     }
 165
 166     /* Start outer loop over neighborlists */
 167     for(iidx=0; iidx<nri; iidx++)
 168     {
 169         /* Load shift vector for this list */
 170         i_shift_offset   = DIM*shiftidx[iidx];
 171
 172         /* Load limits for loop over neighbors */
 173         j_index_start    = jindex[iidx];
 174         j_index_end      = jindex[iidx+1];
 175
 176         /* Get outer coordinate index */
 177         inr              = iinr[iidx];
 178         i_coord_offset   = DIM*inr;
 179
 180         /* Load i particle coords and add shift vector */
 181         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 182                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 183
 184         fix1             = _mm_setzero_ps();
 185         fiy1             = _mm_setzero_ps();
 186         fiz1             = _mm_setzero_ps();
 187         fix2             = _mm_setzero_ps();
 188         fiy2             = _mm_setzero_ps();
 189         fiz2             = _mm_setzero_ps();
 190         fix3             = _mm_setzero_ps();
 191         fiy3             = _mm_setzero_ps();
 192         fiz3             = _mm_setzero_ps();
 193
 194         /* Reset potential sums */
 195         velecsum         = _mm_setzero_ps();
 196
 197         /* Start inner kernel loop */
 198         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 199         {
 200
 201             /* Get j neighbor index, and coordinate index */
 202             jnrA             = jjnr[jidx];
 203             jnrB             = jjnr[jidx+1];
 204             jnrC             = jjnr[jidx+2];
 205             jnrD             = jjnr[jidx+3];
 206             j_coord_offsetA  = DIM*jnrA;
 207             j_coord_offsetB  = DIM*jnrB;
 208             j_coord_offsetC  = DIM*jnrC;
 209             j_coord_offsetD  = DIM*jnrD;
 210
 211             /* load j atom coordinates */
 212             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 213                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 214                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 215
 216             /* Calculate displacement vector */
 217             dx11             = _mm_sub_ps(ix1,jx1);
 218             dy11             = _mm_sub_ps(iy1,jy1);
 219             dz11             = _mm_sub_ps(iz1,jz1);
 220             dx12             = _mm_sub_ps(ix1,jx2);
 221             dy12             = _mm_sub_ps(iy1,jy2);
 222             dz12             = _mm_sub_ps(iz1,jz2);
 223             dx13             = _mm_sub_ps(ix1,jx3);
 224             dy13             = _mm_sub_ps(iy1,jy3);
 225             dz13             = _mm_sub_ps(iz1,jz3);
 226             dx21             = _mm_sub_ps(ix2,jx1);
 227             dy21             = _mm_sub_ps(iy2,jy1);
 228             dz21             = _mm_sub_ps(iz2,jz1);
 229             dx22             = _mm_sub_ps(ix2,jx2);
 230             dy22             = _mm_sub_ps(iy2,jy2);
 231             dz22             = _mm_sub_ps(iz2,jz2);
 232             dx23             = _mm_sub_ps(ix2,jx3);
 233             dy23             = _mm_sub_ps(iy2,jy3);
 234             dz23             = _mm_sub_ps(iz2,jz3);
 235             dx31             = _mm_sub_ps(ix3,jx1);
 236             dy31             = _mm_sub_ps(iy3,jy1);
 237             dz31             = _mm_sub_ps(iz3,jz1);
 238             dx32             = _mm_sub_ps(ix3,jx2);
 239             dy32             = _mm_sub_ps(iy3,jy2);
 240             dz32             = _mm_sub_ps(iz3,jz2);
 241             dx33             = _mm_sub_ps(ix3,jx3);
 242             dy33             = _mm_sub_ps(iy3,jy3);
 243             dz33             = _mm_sub_ps(iz3,jz3);
 244
 245             /* Calculate squared distance and things based on it */
 246             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 247             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 248             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 249             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 250             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 251             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 252             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 253             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 254             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 255
 256             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 257             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 258             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 259             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 260             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 261             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 262             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 263             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 264             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 265
 266             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 267             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 268             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 269             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 270             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 271             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 272             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 273             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 274             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 275
 276             fjx1             = _mm_setzero_ps();
 277             fjy1             = _mm_setzero_ps();
 278             fjz1             = _mm_setzero_ps();
 279             fjx2             = _mm_setzero_ps();
 280             fjy2             = _mm_setzero_ps();
 281             fjz2             = _mm_setzero_ps();
 282             fjx3             = _mm_setzero_ps();
 283             fjy3             = _mm_setzero_ps();
 284             fjz3             = _mm_setzero_ps();
 285
 286             /**************************
 287              * CALCULATE INTERACTIONS *
 288              **************************/
 289
 290             if (gmx_mm_any_lt(rsq11,rcutoff2))
 291             {
 292
 293             /* REACTION-FIELD ELECTROSTATICS */
 294             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 295             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 296
 297             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 298
 299             /* Update potential sum for this i atom from the interaction with this j atom. */
 300             velec            = _mm_and_ps(velec,cutoff_mask);
 301             velecsum         = _mm_add_ps(velecsum,velec);
 302
 303             fscal            = felec;
 304
 305             fscal            = _mm_and_ps(fscal,cutoff_mask);
 306
 307             /* Calculate temporary vectorial force */
 308             tx               = _mm_mul_ps(fscal,dx11);
 309             ty               = _mm_mul_ps(fscal,dy11);
 310             tz               = _mm_mul_ps(fscal,dz11);
 311
 312             /* Update vectorial force */
 313             fix1             = _mm_add_ps(fix1,tx);
 314             fiy1             = _mm_add_ps(fiy1,ty);
 315             fiz1             = _mm_add_ps(fiz1,tz);
 316
 317             fjx1             = _mm_add_ps(fjx1,tx);
 318             fjy1             = _mm_add_ps(fjy1,ty);
 319             fjz1             = _mm_add_ps(fjz1,tz);
 320
 321             }
 322
 323             /**************************
 324              * CALCULATE INTERACTIONS *
 325              **************************/
 326
 327             if (gmx_mm_any_lt(rsq12,rcutoff2))
 328             {
 329
 330             /* REACTION-FIELD ELECTROSTATICS */
 331             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 332             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 333
 334             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 335
 336             /* Update potential sum for this i atom from the interaction with this j atom. */
 337             velec            = _mm_and_ps(velec,cutoff_mask);
 338             velecsum         = _mm_add_ps(velecsum,velec);
 339
 340             fscal            = felec;
 341
 342             fscal            = _mm_and_ps(fscal,cutoff_mask);
 343
 344             /* Calculate temporary vectorial force */
 345             tx               = _mm_mul_ps(fscal,dx12);
 346             ty               = _mm_mul_ps(fscal,dy12);
 347             tz               = _mm_mul_ps(fscal,dz12);
 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             fjx2             = _mm_add_ps(fjx2,tx);
 355             fjy2             = _mm_add_ps(fjy2,ty);
 356             fjz2             = _mm_add_ps(fjz2,tz);
 357
 358             }
 359
 360             /**************************
 361              * CALCULATE INTERACTIONS *
 362              **************************/
 363
 364             if (gmx_mm_any_lt(rsq13,rcutoff2))
 365             {
 366
 367             /* REACTION-FIELD ELECTROSTATICS */
 368             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_add_ps(rinv13,_mm_mul_ps(krf,rsq13)),crf));
 369             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
 370
 371             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
 372
 373             /* Update potential sum for this i atom from the interaction with this j atom. */
 374             velec            = _mm_and_ps(velec,cutoff_mask);
 375             velecsum         = _mm_add_ps(velecsum,velec);
 376
 377             fscal            = felec;
 378
 379             fscal            = _mm_and_ps(fscal,cutoff_mask);
 380
 381             /* Calculate temporary vectorial force */
 382             tx               = _mm_mul_ps(fscal,dx13);
 383             ty               = _mm_mul_ps(fscal,dy13);
 384             tz               = _mm_mul_ps(fscal,dz13);
 385
 386             /* Update vectorial force */
 387             fix1             = _mm_add_ps(fix1,tx);
 388             fiy1             = _mm_add_ps(fiy1,ty);
 389             fiz1             = _mm_add_ps(fiz1,tz);
 390
 391             fjx3             = _mm_add_ps(fjx3,tx);
 392             fjy3             = _mm_add_ps(fjy3,ty);
 393             fjz3             = _mm_add_ps(fjz3,tz);
 394
 395             }
 396
 397             /**************************
 398              * CALCULATE INTERACTIONS *
 399              **************************/
 400
 401             if (gmx_mm_any_lt(rsq21,rcutoff2))
 402             {
 403
 404             /* REACTION-FIELD ELECTROSTATICS */
 405             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 406             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 407
 408             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 409
 410             /* Update potential sum for this i atom from the interaction with this j atom. */
 411             velec            = _mm_and_ps(velec,cutoff_mask);
 412             velecsum         = _mm_add_ps(velecsum,velec);
 413
 414             fscal            = felec;
 415
 416             fscal            = _mm_and_ps(fscal,cutoff_mask);
 417
 418             /* Calculate temporary vectorial force */
 419             tx               = _mm_mul_ps(fscal,dx21);
 420             ty               = _mm_mul_ps(fscal,dy21);
 421             tz               = _mm_mul_ps(fscal,dz21);
 422
 423             /* Update vectorial force */
 424             fix2             = _mm_add_ps(fix2,tx);
 425             fiy2             = _mm_add_ps(fiy2,ty);
 426             fiz2             = _mm_add_ps(fiz2,tz);
 427
 428             fjx1             = _mm_add_ps(fjx1,tx);
 429             fjy1             = _mm_add_ps(fjy1,ty);
 430             fjz1             = _mm_add_ps(fjz1,tz);
 431
 432             }
 433
 434             /**************************
 435              * CALCULATE INTERACTIONS *
 436              **************************/
 437
 438             if (gmx_mm_any_lt(rsq22,rcutoff2))
 439             {
 440
 441             /* REACTION-FIELD ELECTROSTATICS */
 442             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 443             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 444
 445             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 446
 447             /* Update potential sum for this i atom from the interaction with this j atom. */
 448             velec            = _mm_and_ps(velec,cutoff_mask);
 449             velecsum         = _mm_add_ps(velecsum,velec);
 450
 451             fscal            = felec;
 452
 453             fscal            = _mm_and_ps(fscal,cutoff_mask);
 454
 455             /* Calculate temporary vectorial force */
 456             tx               = _mm_mul_ps(fscal,dx22);
 457             ty               = _mm_mul_ps(fscal,dy22);
 458             tz               = _mm_mul_ps(fscal,dz22);
 459
 460             /* Update vectorial force */
 461             fix2             = _mm_add_ps(fix2,tx);
 462             fiy2             = _mm_add_ps(fiy2,ty);
 463             fiz2             = _mm_add_ps(fiz2,tz);
 464
 465             fjx2             = _mm_add_ps(fjx2,tx);
 466             fjy2             = _mm_add_ps(fjy2,ty);
 467             fjz2             = _mm_add_ps(fjz2,tz);
 468
 469             }
 470
 471             /**************************
 472              * CALCULATE INTERACTIONS *
 473              **************************/
 474
 475             if (gmx_mm_any_lt(rsq23,rcutoff2))
 476             {
 477
 478             /* REACTION-FIELD ELECTROSTATICS */
 479             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_add_ps(rinv23,_mm_mul_ps(krf,rsq23)),crf));
 480             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
 481
 482             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
 483
 484             /* Update potential sum for this i atom from the interaction with this j atom. */
 485             velec            = _mm_and_ps(velec,cutoff_mask);
 486             velecsum         = _mm_add_ps(velecsum,velec);
 487
 488             fscal            = felec;
 489
 490             fscal            = _mm_and_ps(fscal,cutoff_mask);
 491
 492             /* Calculate temporary vectorial force */
 493             tx               = _mm_mul_ps(fscal,dx23);
 494             ty               = _mm_mul_ps(fscal,dy23);
 495             tz               = _mm_mul_ps(fscal,dz23);
 496
 497             /* Update vectorial force */
 498             fix2             = _mm_add_ps(fix2,tx);
 499             fiy2             = _mm_add_ps(fiy2,ty);
 500             fiz2             = _mm_add_ps(fiz2,tz);
 501
 502             fjx3             = _mm_add_ps(fjx3,tx);
 503             fjy3             = _mm_add_ps(fjy3,ty);
 504             fjz3             = _mm_add_ps(fjz3,tz);
 505
 506             }
 507
 508             /**************************
 509              * CALCULATE INTERACTIONS *
 510              **************************/
 511
 512             if (gmx_mm_any_lt(rsq31,rcutoff2))
 513             {
 514
 515             /* REACTION-FIELD ELECTROSTATICS */
 516             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_add_ps(rinv31,_mm_mul_ps(krf,rsq31)),crf));
 517             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
 518
 519             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
 520
 521             /* Update potential sum for this i atom from the interaction with this j atom. */
 522             velec            = _mm_and_ps(velec,cutoff_mask);
 523             velecsum         = _mm_add_ps(velecsum,velec);
 524
 525             fscal            = felec;
 526
 527             fscal            = _mm_and_ps(fscal,cutoff_mask);
 528
 529             /* Calculate temporary vectorial force */
 530             tx               = _mm_mul_ps(fscal,dx31);
 531             ty               = _mm_mul_ps(fscal,dy31);
 532             tz               = _mm_mul_ps(fscal,dz31);
 533
 534             /* Update vectorial force */
 535             fix3             = _mm_add_ps(fix3,tx);
 536             fiy3             = _mm_add_ps(fiy3,ty);
 537             fiz3             = _mm_add_ps(fiz3,tz);
 538
 539             fjx1             = _mm_add_ps(fjx1,tx);
 540             fjy1             = _mm_add_ps(fjy1,ty);
 541             fjz1             = _mm_add_ps(fjz1,tz);
 542
 543             }
 544
 545             /**************************
 546              * CALCULATE INTERACTIONS *
 547              **************************/
 548
 549             if (gmx_mm_any_lt(rsq32,rcutoff2))
 550             {
 551
 552             /* REACTION-FIELD ELECTROSTATICS */
 553             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_add_ps(rinv32,_mm_mul_ps(krf,rsq32)),crf));
 554             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
 555
 556             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
 557
 558             /* Update potential sum for this i atom from the interaction with this j atom. */
 559             velec            = _mm_and_ps(velec,cutoff_mask);
 560             velecsum         = _mm_add_ps(velecsum,velec);
 561
 562             fscal            = felec;
 563
 564             fscal            = _mm_and_ps(fscal,cutoff_mask);
 565
 566             /* Calculate temporary vectorial force */
 567             tx               = _mm_mul_ps(fscal,dx32);
 568             ty               = _mm_mul_ps(fscal,dy32);
 569             tz               = _mm_mul_ps(fscal,dz32);
 570
 571             /* Update vectorial force */
 572             fix3             = _mm_add_ps(fix3,tx);
 573             fiy3             = _mm_add_ps(fiy3,ty);
 574             fiz3             = _mm_add_ps(fiz3,tz);
 575
 576             fjx2             = _mm_add_ps(fjx2,tx);
 577             fjy2             = _mm_add_ps(fjy2,ty);
 578             fjz2             = _mm_add_ps(fjz2,tz);
 579
 580             }
 581
 582             /**************************
 583              * CALCULATE INTERACTIONS *
 584              **************************/
 585
 586             if (gmx_mm_any_lt(rsq33,rcutoff2))
 587             {
 588
 589             /* REACTION-FIELD ELECTROSTATICS */
 590             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_add_ps(rinv33,_mm_mul_ps(krf,rsq33)),crf));
 591             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
 592
 593             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
 594
 595             /* Update potential sum for this i atom from the interaction with this j atom. */
 596             velec            = _mm_and_ps(velec,cutoff_mask);
 597             velecsum         = _mm_add_ps(velecsum,velec);
 598
 599             fscal            = felec;
 600
 601             fscal            = _mm_and_ps(fscal,cutoff_mask);
 602
 603             /* Calculate temporary vectorial force */
 604             tx               = _mm_mul_ps(fscal,dx33);
 605             ty               = _mm_mul_ps(fscal,dy33);
 606             tz               = _mm_mul_ps(fscal,dz33);
 607
 608             /* Update vectorial force */
 609             fix3             = _mm_add_ps(fix3,tx);
 610             fiy3             = _mm_add_ps(fiy3,ty);
 611             fiz3             = _mm_add_ps(fiz3,tz);
 612
 613             fjx3             = _mm_add_ps(fjx3,tx);
 614             fjy3             = _mm_add_ps(fjy3,ty);
 615             fjz3             = _mm_add_ps(fjz3,tz);
 616
 617             }
 618
 619             fjptrA             = f+j_coord_offsetA;
 620             fjptrB             = f+j_coord_offsetB;
 621             fjptrC             = f+j_coord_offsetC;
 622             fjptrD             = f+j_coord_offsetD;
 623
 624             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
 625                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 626
 627             /* Inner loop uses 324 flops */
 628         }
 629
 630         if(jidx<j_index_end)
 631         {
 632
 633             /* Get j neighbor index, and coordinate index */
 634             jnrlistA         = jjnr[jidx];
 635             jnrlistB         = jjnr[jidx+1];
 636             jnrlistC         = jjnr[jidx+2];
 637             jnrlistD         = jjnr[jidx+3];
 638             /* Sign of each element will be negative for non-real atoms.
 639              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 640              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 641              */
 642             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 643             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 644             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 645             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 646             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 647             j_coord_offsetA  = DIM*jnrA;
 648             j_coord_offsetB  = DIM*jnrB;
 649             j_coord_offsetC  = DIM*jnrC;
 650             j_coord_offsetD  = DIM*jnrD;
 651
 652             /* load j atom coordinates */
 653             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 654                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
 655                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 656
 657             /* Calculate displacement vector */
 658             dx11             = _mm_sub_ps(ix1,jx1);
 659             dy11             = _mm_sub_ps(iy1,jy1);
 660             dz11             = _mm_sub_ps(iz1,jz1);
 661             dx12             = _mm_sub_ps(ix1,jx2);
 662             dy12             = _mm_sub_ps(iy1,jy2);
 663             dz12             = _mm_sub_ps(iz1,jz2);
 664             dx13             = _mm_sub_ps(ix1,jx3);
 665             dy13             = _mm_sub_ps(iy1,jy3);
 666             dz13             = _mm_sub_ps(iz1,jz3);
 667             dx21             = _mm_sub_ps(ix2,jx1);
 668             dy21             = _mm_sub_ps(iy2,jy1);
 669             dz21             = _mm_sub_ps(iz2,jz1);
 670             dx22             = _mm_sub_ps(ix2,jx2);
 671             dy22             = _mm_sub_ps(iy2,jy2);
 672             dz22             = _mm_sub_ps(iz2,jz2);
 673             dx23             = _mm_sub_ps(ix2,jx3);
 674             dy23             = _mm_sub_ps(iy2,jy3);
 675             dz23             = _mm_sub_ps(iz2,jz3);
 676             dx31             = _mm_sub_ps(ix3,jx1);
 677             dy31             = _mm_sub_ps(iy3,jy1);
 678             dz31             = _mm_sub_ps(iz3,jz1);
 679             dx32             = _mm_sub_ps(ix3,jx2);
 680             dy32             = _mm_sub_ps(iy3,jy2);
 681             dz32             = _mm_sub_ps(iz3,jz2);
 682             dx33             = _mm_sub_ps(ix3,jx3);
 683             dy33             = _mm_sub_ps(iy3,jy3);
 684             dz33             = _mm_sub_ps(iz3,jz3);
 685
 686             /* Calculate squared distance and things based on it */
 687             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 688             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 689             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 690             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 691             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 692             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 693             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 694             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 695             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 696
 697             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 698             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 699             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 700             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 701             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 702             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 703             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 704             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 705             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 706
 707             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 708             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 709             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 710             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 711             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 712             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 713             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 714             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 715             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 716
 717             fjx1             = _mm_setzero_ps();
 718             fjy1             = _mm_setzero_ps();
 719             fjz1             = _mm_setzero_ps();
 720             fjx2             = _mm_setzero_ps();
 721             fjy2             = _mm_setzero_ps();
 722             fjz2             = _mm_setzero_ps();
 723             fjx3             = _mm_setzero_ps();
 724             fjy3             = _mm_setzero_ps();
 725             fjz3             = _mm_setzero_ps();
 726
 727             /**************************
 728              * CALCULATE INTERACTIONS *
 729              **************************/
 730
 731             if (gmx_mm_any_lt(rsq11,rcutoff2))
 732             {
 733
 734             /* REACTION-FIELD ELECTROSTATICS */
 735             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 736             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 737
 738             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 739
 740             /* Update potential sum for this i atom from the interaction with this j atom. */
 741             velec            = _mm_and_ps(velec,cutoff_mask);
 742             velec            = _mm_andnot_ps(dummy_mask,velec);
 743             velecsum         = _mm_add_ps(velecsum,velec);
 744
 745             fscal            = felec;
 746
 747             fscal            = _mm_and_ps(fscal,cutoff_mask);
 748
 749             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 750
 751             /* Calculate temporary vectorial force */
 752             tx               = _mm_mul_ps(fscal,dx11);
 753             ty               = _mm_mul_ps(fscal,dy11);
 754             tz               = _mm_mul_ps(fscal,dz11);
 755
 756             /* Update vectorial force */
 757             fix1             = _mm_add_ps(fix1,tx);
 758             fiy1             = _mm_add_ps(fiy1,ty);
 759             fiz1             = _mm_add_ps(fiz1,tz);
 760
 761             fjx1             = _mm_add_ps(fjx1,tx);
 762             fjy1             = _mm_add_ps(fjy1,ty);
 763             fjz1             = _mm_add_ps(fjz1,tz);
 764
 765             }
 766
 767             /**************************
 768              * CALCULATE INTERACTIONS *
 769              **************************/
 770
 771             if (gmx_mm_any_lt(rsq12,rcutoff2))
 772             {
 773
 774             /* REACTION-FIELD ELECTROSTATICS */
 775             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 776             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 777
 778             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 779
 780             /* Update potential sum for this i atom from the interaction with this j atom. */
 781             velec            = _mm_and_ps(velec,cutoff_mask);
 782             velec            = _mm_andnot_ps(dummy_mask,velec);
 783             velecsum         = _mm_add_ps(velecsum,velec);
 784
 785             fscal            = felec;
 786
 787             fscal            = _mm_and_ps(fscal,cutoff_mask);
 788
 789             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 790
 791             /* Calculate temporary vectorial force */
 792             tx               = _mm_mul_ps(fscal,dx12);
 793             ty               = _mm_mul_ps(fscal,dy12);
 794             tz               = _mm_mul_ps(fscal,dz12);
 795
 796             /* Update vectorial force */
 797             fix1             = _mm_add_ps(fix1,tx);
 798             fiy1             = _mm_add_ps(fiy1,ty);
 799             fiz1             = _mm_add_ps(fiz1,tz);
 800
 801             fjx2             = _mm_add_ps(fjx2,tx);
 802             fjy2             = _mm_add_ps(fjy2,ty);
 803             fjz2             = _mm_add_ps(fjz2,tz);
 804
 805             }
 806
 807             /**************************
 808              * CALCULATE INTERACTIONS *
 809              **************************/
 810
 811             if (gmx_mm_any_lt(rsq13,rcutoff2))
 812             {
 813
 814             /* REACTION-FIELD ELECTROSTATICS */
 815             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_add_ps(rinv13,_mm_mul_ps(krf,rsq13)),crf));
 816             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
 817
 818             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
 819
 820             /* Update potential sum for this i atom from the interaction with this j atom. */
 821             velec            = _mm_and_ps(velec,cutoff_mask);
 822             velec            = _mm_andnot_ps(dummy_mask,velec);
 823             velecsum         = _mm_add_ps(velecsum,velec);
 824
 825             fscal            = felec;
 826
 827             fscal            = _mm_and_ps(fscal,cutoff_mask);
 828
 829             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 830
 831             /* Calculate temporary vectorial force */
 832             tx               = _mm_mul_ps(fscal,dx13);
 833             ty               = _mm_mul_ps(fscal,dy13);
 834             tz               = _mm_mul_ps(fscal,dz13);
 835
 836             /* Update vectorial force */
 837             fix1             = _mm_add_ps(fix1,tx);
 838             fiy1             = _mm_add_ps(fiy1,ty);
 839             fiz1             = _mm_add_ps(fiz1,tz);
 840
 841             fjx3             = _mm_add_ps(fjx3,tx);
 842             fjy3             = _mm_add_ps(fjy3,ty);
 843             fjz3             = _mm_add_ps(fjz3,tz);
 844
 845             }
 846
 847             /**************************
 848              * CALCULATE INTERACTIONS *
 849              **************************/
 850
 851             if (gmx_mm_any_lt(rsq21,rcutoff2))
 852             {
 853
 854             /* REACTION-FIELD ELECTROSTATICS */
 855             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 856             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 857
 858             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 859
 860             /* Update potential sum for this i atom from the interaction with this j atom. */
 861             velec            = _mm_and_ps(velec,cutoff_mask);
 862             velec            = _mm_andnot_ps(dummy_mask,velec);
 863             velecsum         = _mm_add_ps(velecsum,velec);
 864
 865             fscal            = felec;
 866
 867             fscal            = _mm_and_ps(fscal,cutoff_mask);
 868
 869             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 870
 871             /* Calculate temporary vectorial force */
 872             tx               = _mm_mul_ps(fscal,dx21);
 873             ty               = _mm_mul_ps(fscal,dy21);
 874             tz               = _mm_mul_ps(fscal,dz21);
 875
 876             /* Update vectorial force */
 877             fix2             = _mm_add_ps(fix2,tx);
 878             fiy2             = _mm_add_ps(fiy2,ty);
 879             fiz2             = _mm_add_ps(fiz2,tz);
 880
 881             fjx1             = _mm_add_ps(fjx1,tx);
 882             fjy1             = _mm_add_ps(fjy1,ty);
 883             fjz1             = _mm_add_ps(fjz1,tz);
 884
 885             }
 886
 887             /**************************
 888              * CALCULATE INTERACTIONS *
 889              **************************/
 890
 891             if (gmx_mm_any_lt(rsq22,rcutoff2))
 892             {
 893
 894             /* REACTION-FIELD ELECTROSTATICS */
 895             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 896             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 897
 898             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 899
 900             /* Update potential sum for this i atom from the interaction with this j atom. */
 901             velec            = _mm_and_ps(velec,cutoff_mask);
 902             velec            = _mm_andnot_ps(dummy_mask,velec);
 903             velecsum         = _mm_add_ps(velecsum,velec);
 904
 905             fscal            = felec;
 906
 907             fscal            = _mm_and_ps(fscal,cutoff_mask);
 908
 909             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 910
 911             /* Calculate temporary vectorial force */
 912             tx               = _mm_mul_ps(fscal,dx22);
 913             ty               = _mm_mul_ps(fscal,dy22);
 914             tz               = _mm_mul_ps(fscal,dz22);
 915
 916             /* Update vectorial force */
 917             fix2             = _mm_add_ps(fix2,tx);
 918             fiy2             = _mm_add_ps(fiy2,ty);
 919             fiz2             = _mm_add_ps(fiz2,tz);
 920
 921             fjx2             = _mm_add_ps(fjx2,tx);
 922             fjy2             = _mm_add_ps(fjy2,ty);
 923             fjz2             = _mm_add_ps(fjz2,tz);
 924
 925             }
 926
 927             /**************************
 928              * CALCULATE INTERACTIONS *
 929              **************************/
 930
 931             if (gmx_mm_any_lt(rsq23,rcutoff2))
 932             {
 933
 934             /* REACTION-FIELD ELECTROSTATICS */
 935             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_add_ps(rinv23,_mm_mul_ps(krf,rsq23)),crf));
 936             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
 937
 938             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
 939
 940             /* Update potential sum for this i atom from the interaction with this j atom. */
 941             velec            = _mm_and_ps(velec,cutoff_mask);
 942             velec            = _mm_andnot_ps(dummy_mask,velec);
 943             velecsum         = _mm_add_ps(velecsum,velec);
 944
 945             fscal            = felec;
 946
 947             fscal            = _mm_and_ps(fscal,cutoff_mask);
 948
 949             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 950
 951             /* Calculate temporary vectorial force */
 952             tx               = _mm_mul_ps(fscal,dx23);
 953             ty               = _mm_mul_ps(fscal,dy23);
 954             tz               = _mm_mul_ps(fscal,dz23);
 955
 956             /* Update vectorial force */
 957             fix2             = _mm_add_ps(fix2,tx);
 958             fiy2             = _mm_add_ps(fiy2,ty);
 959             fiz2             = _mm_add_ps(fiz2,tz);
 960
 961             fjx3             = _mm_add_ps(fjx3,tx);
 962             fjy3             = _mm_add_ps(fjy3,ty);
 963             fjz3             = _mm_add_ps(fjz3,tz);
 964
 965             }
 966
 967             /**************************
 968              * CALCULATE INTERACTIONS *
 969              **************************/
 970
 971             if (gmx_mm_any_lt(rsq31,rcutoff2))
 972             {
 973
 974             /* REACTION-FIELD ELECTROSTATICS */
 975             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_add_ps(rinv31,_mm_mul_ps(krf,rsq31)),crf));
 976             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
 977
 978             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
 979
 980             /* Update potential sum for this i atom from the interaction with this j atom. */
 981             velec            = _mm_and_ps(velec,cutoff_mask);
 982             velec            = _mm_andnot_ps(dummy_mask,velec);
 983             velecsum         = _mm_add_ps(velecsum,velec);
 984
 985             fscal            = felec;
 986
 987             fscal            = _mm_and_ps(fscal,cutoff_mask);
 988
 989             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 990
 991             /* Calculate temporary vectorial force */
 992             tx               = _mm_mul_ps(fscal,dx31);
 993             ty               = _mm_mul_ps(fscal,dy31);
 994             tz               = _mm_mul_ps(fscal,dz31);
 995
 996             /* Update vectorial force */
 997             fix3             = _mm_add_ps(fix3,tx);
 998             fiy3             = _mm_add_ps(fiy3,ty);
 999             fiz3             = _mm_add_ps(fiz3,tz);
1000
1001             fjx1             = _mm_add_ps(fjx1,tx);
1002             fjy1             = _mm_add_ps(fjy1,ty);
1003             fjz1             = _mm_add_ps(fjz1,tz);
1004
1005             }
1006
1007             /**************************
1008              * CALCULATE INTERACTIONS *
1009              **************************/
1010
1011             if (gmx_mm_any_lt(rsq32,rcutoff2))
1012             {
1013
1014             /* REACTION-FIELD ELECTROSTATICS */
1015             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_add_ps(rinv32,_mm_mul_ps(krf,rsq32)),crf));
1016             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1017
1018             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1019
1020             /* Update potential sum for this i atom from the interaction with this j atom. */
1021             velec            = _mm_and_ps(velec,cutoff_mask);
1022             velec            = _mm_andnot_ps(dummy_mask,velec);
1023             velecsum         = _mm_add_ps(velecsum,velec);
1024
1025             fscal            = felec;
1026
1027             fscal            = _mm_and_ps(fscal,cutoff_mask);
1028
1029             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1030
1031             /* Calculate temporary vectorial force */
1032             tx               = _mm_mul_ps(fscal,dx32);
1033             ty               = _mm_mul_ps(fscal,dy32);
1034             tz               = _mm_mul_ps(fscal,dz32);
1035
1036             /* Update vectorial force */
1037             fix3             = _mm_add_ps(fix3,tx);
1038             fiy3             = _mm_add_ps(fiy3,ty);
1039             fiz3             = _mm_add_ps(fiz3,tz);
1040
1041             fjx2             = _mm_add_ps(fjx2,tx);
1042             fjy2             = _mm_add_ps(fjy2,ty);
1043             fjz2             = _mm_add_ps(fjz2,tz);
1044
1045             }
1046
1047             /**************************
1048              * CALCULATE INTERACTIONS *
1049              **************************/
1050
1051             if (gmx_mm_any_lt(rsq33,rcutoff2))
1052             {
1053
1054             /* REACTION-FIELD ELECTROSTATICS */
1055             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_add_ps(rinv33,_mm_mul_ps(krf,rsq33)),crf));
1056             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
1057
1058             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
1059
1060             /* Update potential sum for this i atom from the interaction with this j atom. */
1061             velec            = _mm_and_ps(velec,cutoff_mask);
1062             velec            = _mm_andnot_ps(dummy_mask,velec);
1063             velecsum         = _mm_add_ps(velecsum,velec);
1064
1065             fscal            = felec;
1066
1067             fscal            = _mm_and_ps(fscal,cutoff_mask);
1068
1069             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1070
1071             /* Calculate temporary vectorial force */
1072             tx               = _mm_mul_ps(fscal,dx33);
1073             ty               = _mm_mul_ps(fscal,dy33);
1074             tz               = _mm_mul_ps(fscal,dz33);
1075
1076             /* Update vectorial force */
1077             fix3             = _mm_add_ps(fix3,tx);
1078             fiy3             = _mm_add_ps(fiy3,ty);
1079             fiz3             = _mm_add_ps(fiz3,tz);
1080
1081             fjx3             = _mm_add_ps(fjx3,tx);
1082             fjy3             = _mm_add_ps(fjy3,ty);
1083             fjz3             = _mm_add_ps(fjz3,tz);
1084
1085             }
1086
1087             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1088             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1089             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1090             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1091
1092             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1093                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1094
1095             /* Inner loop uses 324 flops */
1096         }
1097
1098         /* End of innermost loop */
1099
1100         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1101                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1102
1103         ggid                        = gid[iidx];
1104         /* Update potential energies */
1105         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1106
1107         /* Increment number of inner iterations */
1108         inneriter                  += j_index_end - j_index_start;
1109
1110         /* Outer loop uses 19 flops */
1111     }
1112
1113     /* Increment number of outer iterations */
1114     outeriter        += nri;
1115
1116     /* Update outer/inner flops */
1117
1118     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*324);
1119 }
1120 /*
1121  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwNone_GeomW4W4_F_sse4_1_single
1122  * Electrostatics interaction: ReactionField
1123  * VdW interaction:            None
1124  * Geometry:                   Water4-Water4
1125  * Calculate force/pot:        Force
1126  */
1127 void
1128 nb_kernel_ElecRFCut_VdwNone_GeomW4W4_F_sse4_1_single
1129                     (t_nblist                    * gmx_restrict       nlist,
1130                      rvec                        * gmx_restrict          xx,
1131                      rvec                        * gmx_restrict          ff,
1132                      t_forcerec                  * gmx_restrict          fr,
1133                      t_mdatoms                   * gmx_restrict     mdatoms,
1134                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1135                      t_nrnb                      * gmx_restrict        nrnb)
1136 {
1137     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1138      * just 0 for non-waters.
1139      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1140      * jnr indices corresponding to data put in the four positions in the SIMD register.
1141      */
1142     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1143     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1144     int              jnrA,jnrB,jnrC,jnrD;
1145     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1146     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1147     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1148     real             rcutoff_scalar;
1149     real             *shiftvec,*fshift,*x,*f;
1150     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1151     real             scratch[4*DIM];
1152     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1153     int              vdwioffset1;
1154     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1155     int              vdwioffset2;
1156     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1157     int              vdwioffset3;
1158     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1159     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1160     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1161     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1162     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1163     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1164     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1165     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1166     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1167     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1168     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1169     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1170     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1171     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1172     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1173     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1174     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1175     real             *charge;
1176     __m128           dummy_mask,cutoff_mask;
1177     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1178     __m128           one     = _mm_set1_ps(1.0);
1179     __m128           two     = _mm_set1_ps(2.0);
1180     x                = xx[0];
1181     f                = ff[0];
1182
1183     nri              = nlist->nri;
1184     iinr             = nlist->iinr;
1185     jindex           = nlist->jindex;
1186     jjnr             = nlist->jjnr;
1187     shiftidx         = nlist->shift;
1188     gid              = nlist->gid;
1189     shiftvec         = fr->shift_vec[0];
1190     fshift           = fr->fshift[0];
1191     facel            = _mm_set1_ps(fr->epsfac);
1192     charge           = mdatoms->chargeA;
1193     krf              = _mm_set1_ps(fr->ic->k_rf);
1194     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1195     crf              = _mm_set1_ps(fr->ic->c_rf);
1196
1197     /* Setup water-specific parameters */
1198     inr              = nlist->iinr[0];
1199     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1200     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1201     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1202
1203     jq1              = _mm_set1_ps(charge[inr+1]);
1204     jq2              = _mm_set1_ps(charge[inr+2]);
1205     jq3              = _mm_set1_ps(charge[inr+3]);
1206     qq11             = _mm_mul_ps(iq1,jq1);
1207     qq12             = _mm_mul_ps(iq1,jq2);
1208     qq13             = _mm_mul_ps(iq1,jq3);
1209     qq21             = _mm_mul_ps(iq2,jq1);
1210     qq22             = _mm_mul_ps(iq2,jq2);
1211     qq23             = _mm_mul_ps(iq2,jq3);
1212     qq31             = _mm_mul_ps(iq3,jq1);
1213     qq32             = _mm_mul_ps(iq3,jq2);
1214     qq33             = _mm_mul_ps(iq3,jq3);
1215
1216     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1217     rcutoff_scalar   = fr->rcoulomb;
1218     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1219     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1220
1221     /* Avoid stupid compiler warnings */
1222     jnrA = jnrB = jnrC = jnrD = 0;
1223     j_coord_offsetA = 0;
1224     j_coord_offsetB = 0;
1225     j_coord_offsetC = 0;
1226     j_coord_offsetD = 0;
1227
1228     outeriter        = 0;
1229     inneriter        = 0;
1230
1231     for(iidx=0;iidx<4*DIM;iidx++)
1232     {
1233         scratch[iidx] = 0.0;
1234     }
1235
1236     /* Start outer loop over neighborlists */
1237     for(iidx=0; iidx<nri; iidx++)
1238     {
1239         /* Load shift vector for this list */
1240         i_shift_offset   = DIM*shiftidx[iidx];
1241
1242         /* Load limits for loop over neighbors */
1243         j_index_start    = jindex[iidx];
1244         j_index_end      = jindex[iidx+1];
1245
1246         /* Get outer coordinate index */
1247         inr              = iinr[iidx];
1248         i_coord_offset   = DIM*inr;
1249
1250         /* Load i particle coords and add shift vector */
1251         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
1252                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1253
1254         fix1             = _mm_setzero_ps();
1255         fiy1             = _mm_setzero_ps();
1256         fiz1             = _mm_setzero_ps();
1257         fix2             = _mm_setzero_ps();
1258         fiy2             = _mm_setzero_ps();
1259         fiz2             = _mm_setzero_ps();
1260         fix3             = _mm_setzero_ps();
1261         fiy3             = _mm_setzero_ps();
1262         fiz3             = _mm_setzero_ps();
1263
1264         /* Start inner kernel loop */
1265         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1266         {
1267
1268             /* Get j neighbor index, and coordinate index */
1269             jnrA             = jjnr[jidx];
1270             jnrB             = jjnr[jidx+1];
1271             jnrC             = jjnr[jidx+2];
1272             jnrD             = jjnr[jidx+3];
1273             j_coord_offsetA  = DIM*jnrA;
1274             j_coord_offsetB  = DIM*jnrB;
1275             j_coord_offsetC  = DIM*jnrC;
1276             j_coord_offsetD  = DIM*jnrD;
1277
1278             /* load j atom coordinates */
1279             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1280                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1281                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1282
1283             /* Calculate displacement vector */
1284             dx11             = _mm_sub_ps(ix1,jx1);
1285             dy11             = _mm_sub_ps(iy1,jy1);
1286             dz11             = _mm_sub_ps(iz1,jz1);
1287             dx12             = _mm_sub_ps(ix1,jx2);
1288             dy12             = _mm_sub_ps(iy1,jy2);
1289             dz12             = _mm_sub_ps(iz1,jz2);
1290             dx13             = _mm_sub_ps(ix1,jx3);
1291             dy13             = _mm_sub_ps(iy1,jy3);
1292             dz13             = _mm_sub_ps(iz1,jz3);
1293             dx21             = _mm_sub_ps(ix2,jx1);
1294             dy21             = _mm_sub_ps(iy2,jy1);
1295             dz21             = _mm_sub_ps(iz2,jz1);
1296             dx22             = _mm_sub_ps(ix2,jx2);
1297             dy22             = _mm_sub_ps(iy2,jy2);
1298             dz22             = _mm_sub_ps(iz2,jz2);
1299             dx23             = _mm_sub_ps(ix2,jx3);
1300             dy23             = _mm_sub_ps(iy2,jy3);
1301             dz23             = _mm_sub_ps(iz2,jz3);
1302             dx31             = _mm_sub_ps(ix3,jx1);
1303             dy31             = _mm_sub_ps(iy3,jy1);
1304             dz31             = _mm_sub_ps(iz3,jz1);
1305             dx32             = _mm_sub_ps(ix3,jx2);
1306             dy32             = _mm_sub_ps(iy3,jy2);
1307             dz32             = _mm_sub_ps(iz3,jz2);
1308             dx33             = _mm_sub_ps(ix3,jx3);
1309             dy33             = _mm_sub_ps(iy3,jy3);
1310             dz33             = _mm_sub_ps(iz3,jz3);
1311
1312             /* Calculate squared distance and things based on it */
1313             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1314             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1315             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1316             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1317             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1318             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1319             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1320             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1321             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1322
1323             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1324             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1325             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1326             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1327             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1328             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1329             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1330             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1331             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1332
1333             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1334             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1335             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1336             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1337             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1338             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1339             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1340             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1341             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1342
1343             fjx1             = _mm_setzero_ps();
1344             fjy1             = _mm_setzero_ps();
1345             fjz1             = _mm_setzero_ps();
1346             fjx2             = _mm_setzero_ps();
1347             fjy2             = _mm_setzero_ps();
1348             fjz2             = _mm_setzero_ps();
1349             fjx3             = _mm_setzero_ps();
1350             fjy3             = _mm_setzero_ps();
1351             fjz3             = _mm_setzero_ps();
1352
1353             /**************************
1354              * CALCULATE INTERACTIONS *
1355              **************************/
1356
1357             if (gmx_mm_any_lt(rsq11,rcutoff2))
1358             {
1359
1360             /* REACTION-FIELD ELECTROSTATICS */
1361             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1362
1363             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1364
1365             fscal            = felec;
1366
1367             fscal            = _mm_and_ps(fscal,cutoff_mask);
1368
1369             /* Calculate temporary vectorial force */
1370             tx               = _mm_mul_ps(fscal,dx11);
1371             ty               = _mm_mul_ps(fscal,dy11);
1372             tz               = _mm_mul_ps(fscal,dz11);
1373
1374             /* Update vectorial force */
1375             fix1             = _mm_add_ps(fix1,tx);
1376             fiy1             = _mm_add_ps(fiy1,ty);
1377             fiz1             = _mm_add_ps(fiz1,tz);
1378
1379             fjx1             = _mm_add_ps(fjx1,tx);
1380             fjy1             = _mm_add_ps(fjy1,ty);
1381             fjz1             = _mm_add_ps(fjz1,tz);
1382
1383             }
1384
1385             /**************************
1386              * CALCULATE INTERACTIONS *
1387              **************************/
1388
1389             if (gmx_mm_any_lt(rsq12,rcutoff2))
1390             {
1391
1392             /* REACTION-FIELD ELECTROSTATICS */
1393             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1394
1395             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1396
1397             fscal            = felec;
1398
1399             fscal            = _mm_and_ps(fscal,cutoff_mask);
1400
1401             /* Calculate temporary vectorial force */
1402             tx               = _mm_mul_ps(fscal,dx12);
1403             ty               = _mm_mul_ps(fscal,dy12);
1404             tz               = _mm_mul_ps(fscal,dz12);
1405
1406             /* Update vectorial force */
1407             fix1             = _mm_add_ps(fix1,tx);
1408             fiy1             = _mm_add_ps(fiy1,ty);
1409             fiz1             = _mm_add_ps(fiz1,tz);
1410
1411             fjx2             = _mm_add_ps(fjx2,tx);
1412             fjy2             = _mm_add_ps(fjy2,ty);
1413             fjz2             = _mm_add_ps(fjz2,tz);
1414
1415             }
1416
1417             /**************************
1418              * CALCULATE INTERACTIONS *
1419              **************************/
1420
1421             if (gmx_mm_any_lt(rsq13,rcutoff2))
1422             {
1423
1424             /* REACTION-FIELD ELECTROSTATICS */
1425             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
1426
1427             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
1428
1429             fscal            = felec;
1430
1431             fscal            = _mm_and_ps(fscal,cutoff_mask);
1432
1433             /* Calculate temporary vectorial force */
1434             tx               = _mm_mul_ps(fscal,dx13);
1435             ty               = _mm_mul_ps(fscal,dy13);
1436             tz               = _mm_mul_ps(fscal,dz13);
1437
1438             /* Update vectorial force */
1439             fix1             = _mm_add_ps(fix1,tx);
1440             fiy1             = _mm_add_ps(fiy1,ty);
1441             fiz1             = _mm_add_ps(fiz1,tz);
1442
1443             fjx3             = _mm_add_ps(fjx3,tx);
1444             fjy3             = _mm_add_ps(fjy3,ty);
1445             fjz3             = _mm_add_ps(fjz3,tz);
1446
1447             }
1448
1449             /**************************
1450              * CALCULATE INTERACTIONS *
1451              **************************/
1452
1453             if (gmx_mm_any_lt(rsq21,rcutoff2))
1454             {
1455
1456             /* REACTION-FIELD ELECTROSTATICS */
1457             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1458
1459             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1460
1461             fscal            = felec;
1462
1463             fscal            = _mm_and_ps(fscal,cutoff_mask);
1464
1465             /* Calculate temporary vectorial force */
1466             tx               = _mm_mul_ps(fscal,dx21);
1467             ty               = _mm_mul_ps(fscal,dy21);
1468             tz               = _mm_mul_ps(fscal,dz21);
1469
1470             /* Update vectorial force */
1471             fix2             = _mm_add_ps(fix2,tx);
1472             fiy2             = _mm_add_ps(fiy2,ty);
1473             fiz2             = _mm_add_ps(fiz2,tz);
1474
1475             fjx1             = _mm_add_ps(fjx1,tx);
1476             fjy1             = _mm_add_ps(fjy1,ty);
1477             fjz1             = _mm_add_ps(fjz1,tz);
1478
1479             }
1480
1481             /**************************
1482              * CALCULATE INTERACTIONS *
1483              **************************/
1484
1485             if (gmx_mm_any_lt(rsq22,rcutoff2))
1486             {
1487
1488             /* REACTION-FIELD ELECTROSTATICS */
1489             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1490
1491             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1492
1493             fscal            = felec;
1494
1495             fscal            = _mm_and_ps(fscal,cutoff_mask);
1496
1497             /* Calculate temporary vectorial force */
1498             tx               = _mm_mul_ps(fscal,dx22);
1499             ty               = _mm_mul_ps(fscal,dy22);
1500             tz               = _mm_mul_ps(fscal,dz22);
1501
1502             /* Update vectorial force */
1503             fix2             = _mm_add_ps(fix2,tx);
1504             fiy2             = _mm_add_ps(fiy2,ty);
1505             fiz2             = _mm_add_ps(fiz2,tz);
1506
1507             fjx2             = _mm_add_ps(fjx2,tx);
1508             fjy2             = _mm_add_ps(fjy2,ty);
1509             fjz2             = _mm_add_ps(fjz2,tz);
1510
1511             }
1512
1513             /**************************
1514              * CALCULATE INTERACTIONS *
1515              **************************/
1516
1517             if (gmx_mm_any_lt(rsq23,rcutoff2))
1518             {
1519
1520             /* REACTION-FIELD ELECTROSTATICS */
1521             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
1522
1523             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
1524
1525             fscal            = felec;
1526
1527             fscal            = _mm_and_ps(fscal,cutoff_mask);
1528
1529             /* Calculate temporary vectorial force */
1530             tx               = _mm_mul_ps(fscal,dx23);
1531             ty               = _mm_mul_ps(fscal,dy23);
1532             tz               = _mm_mul_ps(fscal,dz23);
1533
1534             /* Update vectorial force */
1535             fix2             = _mm_add_ps(fix2,tx);
1536             fiy2             = _mm_add_ps(fiy2,ty);
1537             fiz2             = _mm_add_ps(fiz2,tz);
1538
1539             fjx3             = _mm_add_ps(fjx3,tx);
1540             fjy3             = _mm_add_ps(fjy3,ty);
1541             fjz3             = _mm_add_ps(fjz3,tz);
1542
1543             }
1544
1545             /**************************
1546              * CALCULATE INTERACTIONS *
1547              **************************/
1548
1549             if (gmx_mm_any_lt(rsq31,rcutoff2))
1550             {
1551
1552             /* REACTION-FIELD ELECTROSTATICS */
1553             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
1554
1555             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
1556
1557             fscal            = felec;
1558
1559             fscal            = _mm_and_ps(fscal,cutoff_mask);
1560
1561             /* Calculate temporary vectorial force */
1562             tx               = _mm_mul_ps(fscal,dx31);
1563             ty               = _mm_mul_ps(fscal,dy31);
1564             tz               = _mm_mul_ps(fscal,dz31);
1565
1566             /* Update vectorial force */
1567             fix3             = _mm_add_ps(fix3,tx);
1568             fiy3             = _mm_add_ps(fiy3,ty);
1569             fiz3             = _mm_add_ps(fiz3,tz);
1570
1571             fjx1             = _mm_add_ps(fjx1,tx);
1572             fjy1             = _mm_add_ps(fjy1,ty);
1573             fjz1             = _mm_add_ps(fjz1,tz);
1574
1575             }
1576
1577             /**************************
1578              * CALCULATE INTERACTIONS *
1579              **************************/
1580
1581             if (gmx_mm_any_lt(rsq32,rcutoff2))
1582             {
1583
1584             /* REACTION-FIELD ELECTROSTATICS */
1585             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1586
1587             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1588
1589             fscal            = felec;
1590
1591             fscal            = _mm_and_ps(fscal,cutoff_mask);
1592
1593             /* Calculate temporary vectorial force */
1594             tx               = _mm_mul_ps(fscal,dx32);
1595             ty               = _mm_mul_ps(fscal,dy32);
1596             tz               = _mm_mul_ps(fscal,dz32);
1597
1598             /* Update vectorial force */
1599             fix3             = _mm_add_ps(fix3,tx);
1600             fiy3             = _mm_add_ps(fiy3,ty);
1601             fiz3             = _mm_add_ps(fiz3,tz);
1602
1603             fjx2             = _mm_add_ps(fjx2,tx);
1604             fjy2             = _mm_add_ps(fjy2,ty);
1605             fjz2             = _mm_add_ps(fjz2,tz);
1606
1607             }
1608
1609             /**************************
1610              * CALCULATE INTERACTIONS *
1611              **************************/
1612
1613             if (gmx_mm_any_lt(rsq33,rcutoff2))
1614             {
1615
1616             /* REACTION-FIELD ELECTROSTATICS */
1617             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
1618
1619             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
1620
1621             fscal            = felec;
1622
1623             fscal            = _mm_and_ps(fscal,cutoff_mask);
1624
1625             /* Calculate temporary vectorial force */
1626             tx               = _mm_mul_ps(fscal,dx33);
1627             ty               = _mm_mul_ps(fscal,dy33);
1628             tz               = _mm_mul_ps(fscal,dz33);
1629
1630             /* Update vectorial force */
1631             fix3             = _mm_add_ps(fix3,tx);
1632             fiy3             = _mm_add_ps(fiy3,ty);
1633             fiz3             = _mm_add_ps(fiz3,tz);
1634
1635             fjx3             = _mm_add_ps(fjx3,tx);
1636             fjy3             = _mm_add_ps(fjy3,ty);
1637             fjz3             = _mm_add_ps(fjz3,tz);
1638
1639             }
1640
1641             fjptrA             = f+j_coord_offsetA;
1642             fjptrB             = f+j_coord_offsetB;
1643             fjptrC             = f+j_coord_offsetC;
1644             fjptrD             = f+j_coord_offsetD;
1645
1646             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
1647                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1648
1649             /* Inner loop uses 270 flops */
1650         }
1651
1652         if(jidx<j_index_end)
1653         {
1654
1655             /* Get j neighbor index, and coordinate index */
1656             jnrlistA         = jjnr[jidx];
1657             jnrlistB         = jjnr[jidx+1];
1658             jnrlistC         = jjnr[jidx+2];
1659             jnrlistD         = jjnr[jidx+3];
1660             /* Sign of each element will be negative for non-real atoms.
1661              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1662              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1663              */
1664             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1665             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1666             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1667             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1668             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1669             j_coord_offsetA  = DIM*jnrA;
1670             j_coord_offsetB  = DIM*jnrB;
1671             j_coord_offsetC  = DIM*jnrC;
1672             j_coord_offsetD  = DIM*jnrD;
1673
1674             /* load j atom coordinates */
1675             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1676                                               x+j_coord_offsetC+DIM,x+j_coord_offsetD+DIM,
1677                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1678
1679             /* Calculate displacement vector */
1680             dx11             = _mm_sub_ps(ix1,jx1);
1681             dy11             = _mm_sub_ps(iy1,jy1);
1682             dz11             = _mm_sub_ps(iz1,jz1);
1683             dx12             = _mm_sub_ps(ix1,jx2);
1684             dy12             = _mm_sub_ps(iy1,jy2);
1685             dz12             = _mm_sub_ps(iz1,jz2);
1686             dx13             = _mm_sub_ps(ix1,jx3);
1687             dy13             = _mm_sub_ps(iy1,jy3);
1688             dz13             = _mm_sub_ps(iz1,jz3);
1689             dx21             = _mm_sub_ps(ix2,jx1);
1690             dy21             = _mm_sub_ps(iy2,jy1);
1691             dz21             = _mm_sub_ps(iz2,jz1);
1692             dx22             = _mm_sub_ps(ix2,jx2);
1693             dy22             = _mm_sub_ps(iy2,jy2);
1694             dz22             = _mm_sub_ps(iz2,jz2);
1695             dx23             = _mm_sub_ps(ix2,jx3);
1696             dy23             = _mm_sub_ps(iy2,jy3);
1697             dz23             = _mm_sub_ps(iz2,jz3);
1698             dx31             = _mm_sub_ps(ix3,jx1);
1699             dy31             = _mm_sub_ps(iy3,jy1);
1700             dz31             = _mm_sub_ps(iz3,jz1);
1701             dx32             = _mm_sub_ps(ix3,jx2);
1702             dy32             = _mm_sub_ps(iy3,jy2);
1703             dz32             = _mm_sub_ps(iz3,jz2);
1704             dx33             = _mm_sub_ps(ix3,jx3);
1705             dy33             = _mm_sub_ps(iy3,jy3);
1706             dz33             = _mm_sub_ps(iz3,jz3);
1707
1708             /* Calculate squared distance and things based on it */
1709             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1710             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1711             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1712             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1713             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1714             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1715             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1716             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1717             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1718
1719             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1720             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1721             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1722             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1723             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1724             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1725             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1726             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1727             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1728
1729             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1730             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1731             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1732             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1733             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1734             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1735             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1736             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1737             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1738
1739             fjx1             = _mm_setzero_ps();
1740             fjy1             = _mm_setzero_ps();
1741             fjz1             = _mm_setzero_ps();
1742             fjx2             = _mm_setzero_ps();
1743             fjy2             = _mm_setzero_ps();
1744             fjz2             = _mm_setzero_ps();
1745             fjx3             = _mm_setzero_ps();
1746             fjy3             = _mm_setzero_ps();
1747             fjz3             = _mm_setzero_ps();
1748
1749             /**************************
1750              * CALCULATE INTERACTIONS *
1751              **************************/
1752
1753             if (gmx_mm_any_lt(rsq11,rcutoff2))
1754             {
1755
1756             /* REACTION-FIELD ELECTROSTATICS */
1757             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1758
1759             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1760
1761             fscal            = felec;
1762
1763             fscal            = _mm_and_ps(fscal,cutoff_mask);
1764
1765             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1766
1767             /* Calculate temporary vectorial force */
1768             tx               = _mm_mul_ps(fscal,dx11);
1769             ty               = _mm_mul_ps(fscal,dy11);
1770             tz               = _mm_mul_ps(fscal,dz11);
1771
1772             /* Update vectorial force */
1773             fix1             = _mm_add_ps(fix1,tx);
1774             fiy1             = _mm_add_ps(fiy1,ty);
1775             fiz1             = _mm_add_ps(fiz1,tz);
1776
1777             fjx1             = _mm_add_ps(fjx1,tx);
1778             fjy1             = _mm_add_ps(fjy1,ty);
1779             fjz1             = _mm_add_ps(fjz1,tz);
1780
1781             }
1782
1783             /**************************
1784              * CALCULATE INTERACTIONS *
1785              **************************/
1786
1787             if (gmx_mm_any_lt(rsq12,rcutoff2))
1788             {
1789
1790             /* REACTION-FIELD ELECTROSTATICS */
1791             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1792
1793             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1794
1795             fscal            = felec;
1796
1797             fscal            = _mm_and_ps(fscal,cutoff_mask);
1798
1799             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1800
1801             /* Calculate temporary vectorial force */
1802             tx               = _mm_mul_ps(fscal,dx12);
1803             ty               = _mm_mul_ps(fscal,dy12);
1804             tz               = _mm_mul_ps(fscal,dz12);
1805
1806             /* Update vectorial force */
1807             fix1             = _mm_add_ps(fix1,tx);
1808             fiy1             = _mm_add_ps(fiy1,ty);
1809             fiz1             = _mm_add_ps(fiz1,tz);
1810
1811             fjx2             = _mm_add_ps(fjx2,tx);
1812             fjy2             = _mm_add_ps(fjy2,ty);
1813             fjz2             = _mm_add_ps(fjz2,tz);
1814
1815             }
1816
1817             /**************************
1818              * CALCULATE INTERACTIONS *
1819              **************************/
1820
1821             if (gmx_mm_any_lt(rsq13,rcutoff2))
1822             {
1823
1824             /* REACTION-FIELD ELECTROSTATICS */
1825             felec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_mul_ps(rinv13,rinvsq13),krf2));
1826
1827             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
1828
1829             fscal            = felec;
1830
1831             fscal            = _mm_and_ps(fscal,cutoff_mask);
1832
1833             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1834
1835             /* Calculate temporary vectorial force */
1836             tx               = _mm_mul_ps(fscal,dx13);
1837             ty               = _mm_mul_ps(fscal,dy13);
1838             tz               = _mm_mul_ps(fscal,dz13);
1839
1840             /* Update vectorial force */
1841             fix1             = _mm_add_ps(fix1,tx);
1842             fiy1             = _mm_add_ps(fiy1,ty);
1843             fiz1             = _mm_add_ps(fiz1,tz);
1844
1845             fjx3             = _mm_add_ps(fjx3,tx);
1846             fjy3             = _mm_add_ps(fjy3,ty);
1847             fjz3             = _mm_add_ps(fjz3,tz);
1848
1849             }
1850
1851             /**************************
1852              * CALCULATE INTERACTIONS *
1853              **************************/
1854
1855             if (gmx_mm_any_lt(rsq21,rcutoff2))
1856             {
1857
1858             /* REACTION-FIELD ELECTROSTATICS */
1859             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1860
1861             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1862
1863             fscal            = felec;
1864
1865             fscal            = _mm_and_ps(fscal,cutoff_mask);
1866
1867             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1868
1869             /* Calculate temporary vectorial force */
1870             tx               = _mm_mul_ps(fscal,dx21);
1871             ty               = _mm_mul_ps(fscal,dy21);
1872             tz               = _mm_mul_ps(fscal,dz21);
1873
1874             /* Update vectorial force */
1875             fix2             = _mm_add_ps(fix2,tx);
1876             fiy2             = _mm_add_ps(fiy2,ty);
1877             fiz2             = _mm_add_ps(fiz2,tz);
1878
1879             fjx1             = _mm_add_ps(fjx1,tx);
1880             fjy1             = _mm_add_ps(fjy1,ty);
1881             fjz1             = _mm_add_ps(fjz1,tz);
1882
1883             }
1884
1885             /**************************
1886              * CALCULATE INTERACTIONS *
1887              **************************/
1888
1889             if (gmx_mm_any_lt(rsq22,rcutoff2))
1890             {
1891
1892             /* REACTION-FIELD ELECTROSTATICS */
1893             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1894
1895             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1896
1897             fscal            = felec;
1898
1899             fscal            = _mm_and_ps(fscal,cutoff_mask);
1900
1901             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1902
1903             /* Calculate temporary vectorial force */
1904             tx               = _mm_mul_ps(fscal,dx22);
1905             ty               = _mm_mul_ps(fscal,dy22);
1906             tz               = _mm_mul_ps(fscal,dz22);
1907
1908             /* Update vectorial force */
1909             fix2             = _mm_add_ps(fix2,tx);
1910             fiy2             = _mm_add_ps(fiy2,ty);
1911             fiz2             = _mm_add_ps(fiz2,tz);
1912
1913             fjx2             = _mm_add_ps(fjx2,tx);
1914             fjy2             = _mm_add_ps(fjy2,ty);
1915             fjz2             = _mm_add_ps(fjz2,tz);
1916
1917             }
1918
1919             /**************************
1920              * CALCULATE INTERACTIONS *
1921              **************************/
1922
1923             if (gmx_mm_any_lt(rsq23,rcutoff2))
1924             {
1925
1926             /* REACTION-FIELD ELECTROSTATICS */
1927             felec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_mul_ps(rinv23,rinvsq23),krf2));
1928
1929             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
1930
1931             fscal            = felec;
1932
1933             fscal            = _mm_and_ps(fscal,cutoff_mask);
1934
1935             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1936
1937             /* Calculate temporary vectorial force */
1938             tx               = _mm_mul_ps(fscal,dx23);
1939             ty               = _mm_mul_ps(fscal,dy23);
1940             tz               = _mm_mul_ps(fscal,dz23);
1941
1942             /* Update vectorial force */
1943             fix2             = _mm_add_ps(fix2,tx);
1944             fiy2             = _mm_add_ps(fiy2,ty);
1945             fiz2             = _mm_add_ps(fiz2,tz);
1946
1947             fjx3             = _mm_add_ps(fjx3,tx);
1948             fjy3             = _mm_add_ps(fjy3,ty);
1949             fjz3             = _mm_add_ps(fjz3,tz);
1950
1951             }
1952
1953             /**************************
1954              * CALCULATE INTERACTIONS *
1955              **************************/
1956
1957             if (gmx_mm_any_lt(rsq31,rcutoff2))
1958             {
1959
1960             /* REACTION-FIELD ELECTROSTATICS */
1961             felec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_mul_ps(rinv31,rinvsq31),krf2));
1962
1963             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
1964
1965             fscal            = felec;
1966
1967             fscal            = _mm_and_ps(fscal,cutoff_mask);
1968
1969             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1970
1971             /* Calculate temporary vectorial force */
1972             tx               = _mm_mul_ps(fscal,dx31);
1973             ty               = _mm_mul_ps(fscal,dy31);
1974             tz               = _mm_mul_ps(fscal,dz31);
1975
1976             /* Update vectorial force */
1977             fix3             = _mm_add_ps(fix3,tx);
1978             fiy3             = _mm_add_ps(fiy3,ty);
1979             fiz3             = _mm_add_ps(fiz3,tz);
1980
1981             fjx1             = _mm_add_ps(fjx1,tx);
1982             fjy1             = _mm_add_ps(fjy1,ty);
1983             fjz1             = _mm_add_ps(fjz1,tz);
1984
1985             }
1986
1987             /**************************
1988              * CALCULATE INTERACTIONS *
1989              **************************/
1990
1991             if (gmx_mm_any_lt(rsq32,rcutoff2))
1992             {
1993
1994             /* REACTION-FIELD ELECTROSTATICS */
1995             felec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_mul_ps(rinv32,rinvsq32),krf2));
1996
1997             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1998
1999             fscal            = felec;
2000
2001             fscal            = _mm_and_ps(fscal,cutoff_mask);
2002
2003             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2004
2005             /* Calculate temporary vectorial force */
2006             tx               = _mm_mul_ps(fscal,dx32);
2007             ty               = _mm_mul_ps(fscal,dy32);
2008             tz               = _mm_mul_ps(fscal,dz32);
2009
2010             /* Update vectorial force */
2011             fix3             = _mm_add_ps(fix3,tx);
2012             fiy3             = _mm_add_ps(fiy3,ty);
2013             fiz3             = _mm_add_ps(fiz3,tz);
2014
2015             fjx2             = _mm_add_ps(fjx2,tx);
2016             fjy2             = _mm_add_ps(fjy2,ty);
2017             fjz2             = _mm_add_ps(fjz2,tz);
2018
2019             }
2020
2021             /**************************
2022              * CALCULATE INTERACTIONS *
2023              **************************/
2024
2025             if (gmx_mm_any_lt(rsq33,rcutoff2))
2026             {
2027
2028             /* REACTION-FIELD ELECTROSTATICS */
2029             felec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_mul_ps(rinv33,rinvsq33),krf2));
2030
2031             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
2032
2033             fscal            = felec;
2034
2035             fscal            = _mm_and_ps(fscal,cutoff_mask);
2036
2037             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2038
2039             /* Calculate temporary vectorial force */
2040             tx               = _mm_mul_ps(fscal,dx33);
2041             ty               = _mm_mul_ps(fscal,dy33);
2042             tz               = _mm_mul_ps(fscal,dz33);
2043
2044             /* Update vectorial force */
2045             fix3             = _mm_add_ps(fix3,tx);
2046             fiy3             = _mm_add_ps(fiy3,ty);
2047             fiz3             = _mm_add_ps(fiz3,tz);
2048
2049             fjx3             = _mm_add_ps(fjx3,tx);
2050             fjy3             = _mm_add_ps(fjy3,ty);
2051             fjz3             = _mm_add_ps(fjz3,tz);
2052
2053             }
2054
2055             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2056             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2057             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2058             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2059
2060             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA+DIM,fjptrB+DIM,fjptrC+DIM,fjptrD+DIM,
2061                                                    fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2062
2063             /* Inner loop uses 270 flops */
2064         }
2065
2066         /* End of innermost loop */
2067
2068         gmx_mm_update_iforce_3atom_swizzle_ps(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2069                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
2070
2071         /* Increment number of inner iterations */
2072         inneriter                  += j_index_end - j_index_start;
2073
2074         /* Outer loop uses 18 flops */
2075     }
2076
2077     /* Increment number of outer iterations */
2078     outeriter        += nri;
2079
2080     /* Update outer/inner flops */
2081
2082     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*270);
2083 }