src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecRFCut_VdwNone_GeomW3W3_sse2_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS sse2_single kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_sse2_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwNone_GeomW3W3_VF_sse2_single
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            None
  53  * Geometry:                   Water3-Water3
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRFCut_VdwNone_GeomW3W3_VF_sse2_single
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  75     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  80     real             scratch[4*DIM];
  81     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwioffset1;
  85     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  86     int              vdwioffset2;
  87     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  88     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  89     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  91     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  92     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  93     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  94     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  95     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  96     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  97     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  98     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  99     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 100     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 101     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 102     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 103     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 104     real             *charge;
 105     __m128           dummy_mask,cutoff_mask;
 106     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 107     __m128           one     = _mm_set1_ps(1.0);
 108     __m128           two     = _mm_set1_ps(2.0);
 109     x                = xx[0];
 110     f                = ff[0];
 111
 112     nri              = nlist->nri;
 113     iinr             = nlist->iinr;
 114     jindex           = nlist->jindex;
 115     jjnr             = nlist->jjnr;
 116     shiftidx         = nlist->shift;
 117     gid              = nlist->gid;
 118     shiftvec         = fr->shift_vec[0];
 119     fshift           = fr->fshift[0];
 120     facel            = _mm_set1_ps(fr->ic->epsfac);
 121     charge           = mdatoms->chargeA;
 122     krf              = _mm_set1_ps(fr->ic->k_rf);
 123     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 124     crf              = _mm_set1_ps(fr->ic->c_rf);
 125
 126     /* Setup water-specific parameters */
 127     inr              = nlist->iinr[0];
 128     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 129     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 130     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 131
 132     jq0              = _mm_set1_ps(charge[inr+0]);
 133     jq1              = _mm_set1_ps(charge[inr+1]);
 134     jq2              = _mm_set1_ps(charge[inr+2]);
 135     qq00             = _mm_mul_ps(iq0,jq0);
 136     qq01             = _mm_mul_ps(iq0,jq1);
 137     qq02             = _mm_mul_ps(iq0,jq2);
 138     qq10             = _mm_mul_ps(iq1,jq0);
 139     qq11             = _mm_mul_ps(iq1,jq1);
 140     qq12             = _mm_mul_ps(iq1,jq2);
 141     qq20             = _mm_mul_ps(iq2,jq0);
 142     qq21             = _mm_mul_ps(iq2,jq1);
 143     qq22             = _mm_mul_ps(iq2,jq2);
 144
 145     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 146     rcutoff_scalar   = fr->ic->rcoulomb;
 147     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 148     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 149
 150     /* Avoid stupid compiler warnings */
 151     jnrA = jnrB = jnrC = jnrD = 0;
 152     j_coord_offsetA = 0;
 153     j_coord_offsetB = 0;
 154     j_coord_offsetC = 0;
 155     j_coord_offsetD = 0;
 156
 157     outeriter        = 0;
 158     inneriter        = 0;
 159
 160     for(iidx=0;iidx<4*DIM;iidx++)
 161     {
 162         scratch[iidx] = 0.0;
 163     }
 164
 165     /* Start outer loop over neighborlists */
 166     for(iidx=0; iidx<nri; iidx++)
 167     {
 168         /* Load shift vector for this list */
 169         i_shift_offset   = DIM*shiftidx[iidx];
 170
 171         /* Load limits for loop over neighbors */
 172         j_index_start    = jindex[iidx];
 173         j_index_end      = jindex[iidx+1];
 174
 175         /* Get outer coordinate index */
 176         inr              = iinr[iidx];
 177         i_coord_offset   = DIM*inr;
 178
 179         /* Load i particle coords and add shift vector */
 180         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 181                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 182
 183         fix0             = _mm_setzero_ps();
 184         fiy0             = _mm_setzero_ps();
 185         fiz0             = _mm_setzero_ps();
 186         fix1             = _mm_setzero_ps();
 187         fiy1             = _mm_setzero_ps();
 188         fiz1             = _mm_setzero_ps();
 189         fix2             = _mm_setzero_ps();
 190         fiy2             = _mm_setzero_ps();
 191         fiz2             = _mm_setzero_ps();
 192
 193         /* Reset potential sums */
 194         velecsum         = _mm_setzero_ps();
 195
 196         /* Start inner kernel loop */
 197         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 198         {
 199
 200             /* Get j neighbor index, and coordinate index */
 201             jnrA             = jjnr[jidx];
 202             jnrB             = jjnr[jidx+1];
 203             jnrC             = jjnr[jidx+2];
 204             jnrD             = jjnr[jidx+3];
 205             j_coord_offsetA  = DIM*jnrA;
 206             j_coord_offsetB  = DIM*jnrB;
 207             j_coord_offsetC  = DIM*jnrC;
 208             j_coord_offsetD  = DIM*jnrD;
 209
 210             /* load j atom coordinates */
 211             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 212                                               x+j_coord_offsetC,x+j_coord_offsetD,
 213                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 214
 215             /* Calculate displacement vector */
 216             dx00             = _mm_sub_ps(ix0,jx0);
 217             dy00             = _mm_sub_ps(iy0,jy0);
 218             dz00             = _mm_sub_ps(iz0,jz0);
 219             dx01             = _mm_sub_ps(ix0,jx1);
 220             dy01             = _mm_sub_ps(iy0,jy1);
 221             dz01             = _mm_sub_ps(iz0,jz1);
 222             dx02             = _mm_sub_ps(ix0,jx2);
 223             dy02             = _mm_sub_ps(iy0,jy2);
 224             dz02             = _mm_sub_ps(iz0,jz2);
 225             dx10             = _mm_sub_ps(ix1,jx0);
 226             dy10             = _mm_sub_ps(iy1,jy0);
 227             dz10             = _mm_sub_ps(iz1,jz0);
 228             dx11             = _mm_sub_ps(ix1,jx1);
 229             dy11             = _mm_sub_ps(iy1,jy1);
 230             dz11             = _mm_sub_ps(iz1,jz1);
 231             dx12             = _mm_sub_ps(ix1,jx2);
 232             dy12             = _mm_sub_ps(iy1,jy2);
 233             dz12             = _mm_sub_ps(iz1,jz2);
 234             dx20             = _mm_sub_ps(ix2,jx0);
 235             dy20             = _mm_sub_ps(iy2,jy0);
 236             dz20             = _mm_sub_ps(iz2,jz0);
 237             dx21             = _mm_sub_ps(ix2,jx1);
 238             dy21             = _mm_sub_ps(iy2,jy1);
 239             dz21             = _mm_sub_ps(iz2,jz1);
 240             dx22             = _mm_sub_ps(ix2,jx2);
 241             dy22             = _mm_sub_ps(iy2,jy2);
 242             dz22             = _mm_sub_ps(iz2,jz2);
 243
 244             /* Calculate squared distance and things based on it */
 245             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 246             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 247             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 248             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 249             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 250             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 251             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 252             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 253             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 254
 255             rinv00           = sse2_invsqrt_f(rsq00);
 256             rinv01           = sse2_invsqrt_f(rsq01);
 257             rinv02           = sse2_invsqrt_f(rsq02);
 258             rinv10           = sse2_invsqrt_f(rsq10);
 259             rinv11           = sse2_invsqrt_f(rsq11);
 260             rinv12           = sse2_invsqrt_f(rsq12);
 261             rinv20           = sse2_invsqrt_f(rsq20);
 262             rinv21           = sse2_invsqrt_f(rsq21);
 263             rinv22           = sse2_invsqrt_f(rsq22);
 264
 265             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 266             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 267             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 268             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 269             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 270             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 271             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 272             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 273             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 274
 275             fjx0             = _mm_setzero_ps();
 276             fjy0             = _mm_setzero_ps();
 277             fjz0             = _mm_setzero_ps();
 278             fjx1             = _mm_setzero_ps();
 279             fjy1             = _mm_setzero_ps();
 280             fjz1             = _mm_setzero_ps();
 281             fjx2             = _mm_setzero_ps();
 282             fjy2             = _mm_setzero_ps();
 283             fjz2             = _mm_setzero_ps();
 284
 285             /**************************
 286              * CALCULATE INTERACTIONS *
 287              **************************/
 288
 289             if (gmx_mm_any_lt(rsq00,rcutoff2))
 290             {
 291
 292             /* REACTION-FIELD ELECTROSTATICS */
 293             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_add_ps(rinv00,_mm_mul_ps(krf,rsq00)),crf));
 294             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
 295
 296             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 297
 298             /* Update potential sum for this i atom from the interaction with this j atom. */
 299             velec            = _mm_and_ps(velec,cutoff_mask);
 300             velecsum         = _mm_add_ps(velecsum,velec);
 301
 302             fscal            = felec;
 303
 304             fscal            = _mm_and_ps(fscal,cutoff_mask);
 305
 306             /* Calculate temporary vectorial force */
 307             tx               = _mm_mul_ps(fscal,dx00);
 308             ty               = _mm_mul_ps(fscal,dy00);
 309             tz               = _mm_mul_ps(fscal,dz00);
 310
 311             /* Update vectorial force */
 312             fix0             = _mm_add_ps(fix0,tx);
 313             fiy0             = _mm_add_ps(fiy0,ty);
 314             fiz0             = _mm_add_ps(fiz0,tz);
 315
 316             fjx0             = _mm_add_ps(fjx0,tx);
 317             fjy0             = _mm_add_ps(fjy0,ty);
 318             fjz0             = _mm_add_ps(fjz0,tz);
 319
 320             }
 321
 322             /**************************
 323              * CALCULATE INTERACTIONS *
 324              **************************/
 325
 326             if (gmx_mm_any_lt(rsq01,rcutoff2))
 327             {
 328
 329             /* REACTION-FIELD ELECTROSTATICS */
 330             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_add_ps(rinv01,_mm_mul_ps(krf,rsq01)),crf));
 331             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
 332
 333             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 334
 335             /* Update potential sum for this i atom from the interaction with this j atom. */
 336             velec            = _mm_and_ps(velec,cutoff_mask);
 337             velecsum         = _mm_add_ps(velecsum,velec);
 338
 339             fscal            = felec;
 340
 341             fscal            = _mm_and_ps(fscal,cutoff_mask);
 342
 343             /* Calculate temporary vectorial force */
 344             tx               = _mm_mul_ps(fscal,dx01);
 345             ty               = _mm_mul_ps(fscal,dy01);
 346             tz               = _mm_mul_ps(fscal,dz01);
 347
 348             /* Update vectorial force */
 349             fix0             = _mm_add_ps(fix0,tx);
 350             fiy0             = _mm_add_ps(fiy0,ty);
 351             fiz0             = _mm_add_ps(fiz0,tz);
 352
 353             fjx1             = _mm_add_ps(fjx1,tx);
 354             fjy1             = _mm_add_ps(fjy1,ty);
 355             fjz1             = _mm_add_ps(fjz1,tz);
 356
 357             }
 358
 359             /**************************
 360              * CALCULATE INTERACTIONS *
 361              **************************/
 362
 363             if (gmx_mm_any_lt(rsq02,rcutoff2))
 364             {
 365
 366             /* REACTION-FIELD ELECTROSTATICS */
 367             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_add_ps(rinv02,_mm_mul_ps(krf,rsq02)),crf));
 368             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
 369
 370             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 371
 372             /* Update potential sum for this i atom from the interaction with this j atom. */
 373             velec            = _mm_and_ps(velec,cutoff_mask);
 374             velecsum         = _mm_add_ps(velecsum,velec);
 375
 376             fscal            = felec;
 377
 378             fscal            = _mm_and_ps(fscal,cutoff_mask);
 379
 380             /* Calculate temporary vectorial force */
 381             tx               = _mm_mul_ps(fscal,dx02);
 382             ty               = _mm_mul_ps(fscal,dy02);
 383             tz               = _mm_mul_ps(fscal,dz02);
 384
 385             /* Update vectorial force */
 386             fix0             = _mm_add_ps(fix0,tx);
 387             fiy0             = _mm_add_ps(fiy0,ty);
 388             fiz0             = _mm_add_ps(fiz0,tz);
 389
 390             fjx2             = _mm_add_ps(fjx2,tx);
 391             fjy2             = _mm_add_ps(fjy2,ty);
 392             fjz2             = _mm_add_ps(fjz2,tz);
 393
 394             }
 395
 396             /**************************
 397              * CALCULATE INTERACTIONS *
 398              **************************/
 399
 400             if (gmx_mm_any_lt(rsq10,rcutoff2))
 401             {
 402
 403             /* REACTION-FIELD ELECTROSTATICS */
 404             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_add_ps(rinv10,_mm_mul_ps(krf,rsq10)),crf));
 405             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
 406
 407             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 408
 409             /* Update potential sum for this i atom from the interaction with this j atom. */
 410             velec            = _mm_and_ps(velec,cutoff_mask);
 411             velecsum         = _mm_add_ps(velecsum,velec);
 412
 413             fscal            = felec;
 414
 415             fscal            = _mm_and_ps(fscal,cutoff_mask);
 416
 417             /* Calculate temporary vectorial force */
 418             tx               = _mm_mul_ps(fscal,dx10);
 419             ty               = _mm_mul_ps(fscal,dy10);
 420             tz               = _mm_mul_ps(fscal,dz10);
 421
 422             /* Update vectorial force */
 423             fix1             = _mm_add_ps(fix1,tx);
 424             fiy1             = _mm_add_ps(fiy1,ty);
 425             fiz1             = _mm_add_ps(fiz1,tz);
 426
 427             fjx0             = _mm_add_ps(fjx0,tx);
 428             fjy0             = _mm_add_ps(fjy0,ty);
 429             fjz0             = _mm_add_ps(fjz0,tz);
 430
 431             }
 432
 433             /**************************
 434              * CALCULATE INTERACTIONS *
 435              **************************/
 436
 437             if (gmx_mm_any_lt(rsq11,rcutoff2))
 438             {
 439
 440             /* REACTION-FIELD ELECTROSTATICS */
 441             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 442             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 443
 444             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 445
 446             /* Update potential sum for this i atom from the interaction with this j atom. */
 447             velec            = _mm_and_ps(velec,cutoff_mask);
 448             velecsum         = _mm_add_ps(velecsum,velec);
 449
 450             fscal            = felec;
 451
 452             fscal            = _mm_and_ps(fscal,cutoff_mask);
 453
 454             /* Calculate temporary vectorial force */
 455             tx               = _mm_mul_ps(fscal,dx11);
 456             ty               = _mm_mul_ps(fscal,dy11);
 457             tz               = _mm_mul_ps(fscal,dz11);
 458
 459             /* Update vectorial force */
 460             fix1             = _mm_add_ps(fix1,tx);
 461             fiy1             = _mm_add_ps(fiy1,ty);
 462             fiz1             = _mm_add_ps(fiz1,tz);
 463
 464             fjx1             = _mm_add_ps(fjx1,tx);
 465             fjy1             = _mm_add_ps(fjy1,ty);
 466             fjz1             = _mm_add_ps(fjz1,tz);
 467
 468             }
 469
 470             /**************************
 471              * CALCULATE INTERACTIONS *
 472              **************************/
 473
 474             if (gmx_mm_any_lt(rsq12,rcutoff2))
 475             {
 476
 477             /* REACTION-FIELD ELECTROSTATICS */
 478             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 479             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 480
 481             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 482
 483             /* Update potential sum for this i atom from the interaction with this j atom. */
 484             velec            = _mm_and_ps(velec,cutoff_mask);
 485             velecsum         = _mm_add_ps(velecsum,velec);
 486
 487             fscal            = felec;
 488
 489             fscal            = _mm_and_ps(fscal,cutoff_mask);
 490
 491             /* Calculate temporary vectorial force */
 492             tx               = _mm_mul_ps(fscal,dx12);
 493             ty               = _mm_mul_ps(fscal,dy12);
 494             tz               = _mm_mul_ps(fscal,dz12);
 495
 496             /* Update vectorial force */
 497             fix1             = _mm_add_ps(fix1,tx);
 498             fiy1             = _mm_add_ps(fiy1,ty);
 499             fiz1             = _mm_add_ps(fiz1,tz);
 500
 501             fjx2             = _mm_add_ps(fjx2,tx);
 502             fjy2             = _mm_add_ps(fjy2,ty);
 503             fjz2             = _mm_add_ps(fjz2,tz);
 504
 505             }
 506
 507             /**************************
 508              * CALCULATE INTERACTIONS *
 509              **************************/
 510
 511             if (gmx_mm_any_lt(rsq20,rcutoff2))
 512             {
 513
 514             /* REACTION-FIELD ELECTROSTATICS */
 515             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_add_ps(rinv20,_mm_mul_ps(krf,rsq20)),crf));
 516             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
 517
 518             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 519
 520             /* Update potential sum for this i atom from the interaction with this j atom. */
 521             velec            = _mm_and_ps(velec,cutoff_mask);
 522             velecsum         = _mm_add_ps(velecsum,velec);
 523
 524             fscal            = felec;
 525
 526             fscal            = _mm_and_ps(fscal,cutoff_mask);
 527
 528             /* Calculate temporary vectorial force */
 529             tx               = _mm_mul_ps(fscal,dx20);
 530             ty               = _mm_mul_ps(fscal,dy20);
 531             tz               = _mm_mul_ps(fscal,dz20);
 532
 533             /* Update vectorial force */
 534             fix2             = _mm_add_ps(fix2,tx);
 535             fiy2             = _mm_add_ps(fiy2,ty);
 536             fiz2             = _mm_add_ps(fiz2,tz);
 537
 538             fjx0             = _mm_add_ps(fjx0,tx);
 539             fjy0             = _mm_add_ps(fjy0,ty);
 540             fjz0             = _mm_add_ps(fjz0,tz);
 541
 542             }
 543
 544             /**************************
 545              * CALCULATE INTERACTIONS *
 546              **************************/
 547
 548             if (gmx_mm_any_lt(rsq21,rcutoff2))
 549             {
 550
 551             /* REACTION-FIELD ELECTROSTATICS */
 552             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 553             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 554
 555             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 556
 557             /* Update potential sum for this i atom from the interaction with this j atom. */
 558             velec            = _mm_and_ps(velec,cutoff_mask);
 559             velecsum         = _mm_add_ps(velecsum,velec);
 560
 561             fscal            = felec;
 562
 563             fscal            = _mm_and_ps(fscal,cutoff_mask);
 564
 565             /* Calculate temporary vectorial force */
 566             tx               = _mm_mul_ps(fscal,dx21);
 567             ty               = _mm_mul_ps(fscal,dy21);
 568             tz               = _mm_mul_ps(fscal,dz21);
 569
 570             /* Update vectorial force */
 571             fix2             = _mm_add_ps(fix2,tx);
 572             fiy2             = _mm_add_ps(fiy2,ty);
 573             fiz2             = _mm_add_ps(fiz2,tz);
 574
 575             fjx1             = _mm_add_ps(fjx1,tx);
 576             fjy1             = _mm_add_ps(fjy1,ty);
 577             fjz1             = _mm_add_ps(fjz1,tz);
 578
 579             }
 580
 581             /**************************
 582              * CALCULATE INTERACTIONS *
 583              **************************/
 584
 585             if (gmx_mm_any_lt(rsq22,rcutoff2))
 586             {
 587
 588             /* REACTION-FIELD ELECTROSTATICS */
 589             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 590             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 591
 592             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 593
 594             /* Update potential sum for this i atom from the interaction with this j atom. */
 595             velec            = _mm_and_ps(velec,cutoff_mask);
 596             velecsum         = _mm_add_ps(velecsum,velec);
 597
 598             fscal            = felec;
 599
 600             fscal            = _mm_and_ps(fscal,cutoff_mask);
 601
 602             /* Calculate temporary vectorial force */
 603             tx               = _mm_mul_ps(fscal,dx22);
 604             ty               = _mm_mul_ps(fscal,dy22);
 605             tz               = _mm_mul_ps(fscal,dz22);
 606
 607             /* Update vectorial force */
 608             fix2             = _mm_add_ps(fix2,tx);
 609             fiy2             = _mm_add_ps(fiy2,ty);
 610             fiz2             = _mm_add_ps(fiz2,tz);
 611
 612             fjx2             = _mm_add_ps(fjx2,tx);
 613             fjy2             = _mm_add_ps(fjy2,ty);
 614             fjz2             = _mm_add_ps(fjz2,tz);
 615
 616             }
 617
 618             fjptrA             = f+j_coord_offsetA;
 619             fjptrB             = f+j_coord_offsetB;
 620             fjptrC             = f+j_coord_offsetC;
 621             fjptrD             = f+j_coord_offsetD;
 622
 623             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 624                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 625
 626             /* Inner loop uses 324 flops */
 627         }
 628
 629         if(jidx<j_index_end)
 630         {
 631
 632             /* Get j neighbor index, and coordinate index */
 633             jnrlistA         = jjnr[jidx];
 634             jnrlistB         = jjnr[jidx+1];
 635             jnrlistC         = jjnr[jidx+2];
 636             jnrlistD         = jjnr[jidx+3];
 637             /* Sign of each element will be negative for non-real atoms.
 638              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 639              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 640              */
 641             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 642             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 643             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 644             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 645             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 646             j_coord_offsetA  = DIM*jnrA;
 647             j_coord_offsetB  = DIM*jnrB;
 648             j_coord_offsetC  = DIM*jnrC;
 649             j_coord_offsetD  = DIM*jnrD;
 650
 651             /* load j atom coordinates */
 652             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 653                                               x+j_coord_offsetC,x+j_coord_offsetD,
 654                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 655
 656             /* Calculate displacement vector */
 657             dx00             = _mm_sub_ps(ix0,jx0);
 658             dy00             = _mm_sub_ps(iy0,jy0);
 659             dz00             = _mm_sub_ps(iz0,jz0);
 660             dx01             = _mm_sub_ps(ix0,jx1);
 661             dy01             = _mm_sub_ps(iy0,jy1);
 662             dz01             = _mm_sub_ps(iz0,jz1);
 663             dx02             = _mm_sub_ps(ix0,jx2);
 664             dy02             = _mm_sub_ps(iy0,jy2);
 665             dz02             = _mm_sub_ps(iz0,jz2);
 666             dx10             = _mm_sub_ps(ix1,jx0);
 667             dy10             = _mm_sub_ps(iy1,jy0);
 668             dz10             = _mm_sub_ps(iz1,jz0);
 669             dx11             = _mm_sub_ps(ix1,jx1);
 670             dy11             = _mm_sub_ps(iy1,jy1);
 671             dz11             = _mm_sub_ps(iz1,jz1);
 672             dx12             = _mm_sub_ps(ix1,jx2);
 673             dy12             = _mm_sub_ps(iy1,jy2);
 674             dz12             = _mm_sub_ps(iz1,jz2);
 675             dx20             = _mm_sub_ps(ix2,jx0);
 676             dy20             = _mm_sub_ps(iy2,jy0);
 677             dz20             = _mm_sub_ps(iz2,jz0);
 678             dx21             = _mm_sub_ps(ix2,jx1);
 679             dy21             = _mm_sub_ps(iy2,jy1);
 680             dz21             = _mm_sub_ps(iz2,jz1);
 681             dx22             = _mm_sub_ps(ix2,jx2);
 682             dy22             = _mm_sub_ps(iy2,jy2);
 683             dz22             = _mm_sub_ps(iz2,jz2);
 684
 685             /* Calculate squared distance and things based on it */
 686             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 687             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 688             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 689             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 690             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 691             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 692             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 693             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 694             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 695
 696             rinv00           = sse2_invsqrt_f(rsq00);
 697             rinv01           = sse2_invsqrt_f(rsq01);
 698             rinv02           = sse2_invsqrt_f(rsq02);
 699             rinv10           = sse2_invsqrt_f(rsq10);
 700             rinv11           = sse2_invsqrt_f(rsq11);
 701             rinv12           = sse2_invsqrt_f(rsq12);
 702             rinv20           = sse2_invsqrt_f(rsq20);
 703             rinv21           = sse2_invsqrt_f(rsq21);
 704             rinv22           = sse2_invsqrt_f(rsq22);
 705
 706             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 707             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 708             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 709             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 710             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 711             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 712             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 713             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 714             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 715
 716             fjx0             = _mm_setzero_ps();
 717             fjy0             = _mm_setzero_ps();
 718             fjz0             = _mm_setzero_ps();
 719             fjx1             = _mm_setzero_ps();
 720             fjy1             = _mm_setzero_ps();
 721             fjz1             = _mm_setzero_ps();
 722             fjx2             = _mm_setzero_ps();
 723             fjy2             = _mm_setzero_ps();
 724             fjz2             = _mm_setzero_ps();
 725
 726             /**************************
 727              * CALCULATE INTERACTIONS *
 728              **************************/
 729
 730             if (gmx_mm_any_lt(rsq00,rcutoff2))
 731             {
 732
 733             /* REACTION-FIELD ELECTROSTATICS */
 734             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_add_ps(rinv00,_mm_mul_ps(krf,rsq00)),crf));
 735             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
 736
 737             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 738
 739             /* Update potential sum for this i atom from the interaction with this j atom. */
 740             velec            = _mm_and_ps(velec,cutoff_mask);
 741             velec            = _mm_andnot_ps(dummy_mask,velec);
 742             velecsum         = _mm_add_ps(velecsum,velec);
 743
 744             fscal            = felec;
 745
 746             fscal            = _mm_and_ps(fscal,cutoff_mask);
 747
 748             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 749
 750             /* Calculate temporary vectorial force */
 751             tx               = _mm_mul_ps(fscal,dx00);
 752             ty               = _mm_mul_ps(fscal,dy00);
 753             tz               = _mm_mul_ps(fscal,dz00);
 754
 755             /* Update vectorial force */
 756             fix0             = _mm_add_ps(fix0,tx);
 757             fiy0             = _mm_add_ps(fiy0,ty);
 758             fiz0             = _mm_add_ps(fiz0,tz);
 759
 760             fjx0             = _mm_add_ps(fjx0,tx);
 761             fjy0             = _mm_add_ps(fjy0,ty);
 762             fjz0             = _mm_add_ps(fjz0,tz);
 763
 764             }
 765
 766             /**************************
 767              * CALCULATE INTERACTIONS *
 768              **************************/
 769
 770             if (gmx_mm_any_lt(rsq01,rcutoff2))
 771             {
 772
 773             /* REACTION-FIELD ELECTROSTATICS */
 774             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_add_ps(rinv01,_mm_mul_ps(krf,rsq01)),crf));
 775             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
 776
 777             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 778
 779             /* Update potential sum for this i atom from the interaction with this j atom. */
 780             velec            = _mm_and_ps(velec,cutoff_mask);
 781             velec            = _mm_andnot_ps(dummy_mask,velec);
 782             velecsum         = _mm_add_ps(velecsum,velec);
 783
 784             fscal            = felec;
 785
 786             fscal            = _mm_and_ps(fscal,cutoff_mask);
 787
 788             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 789
 790             /* Calculate temporary vectorial force */
 791             tx               = _mm_mul_ps(fscal,dx01);
 792             ty               = _mm_mul_ps(fscal,dy01);
 793             tz               = _mm_mul_ps(fscal,dz01);
 794
 795             /* Update vectorial force */
 796             fix0             = _mm_add_ps(fix0,tx);
 797             fiy0             = _mm_add_ps(fiy0,ty);
 798             fiz0             = _mm_add_ps(fiz0,tz);
 799
 800             fjx1             = _mm_add_ps(fjx1,tx);
 801             fjy1             = _mm_add_ps(fjy1,ty);
 802             fjz1             = _mm_add_ps(fjz1,tz);
 803
 804             }
 805
 806             /**************************
 807              * CALCULATE INTERACTIONS *
 808              **************************/
 809
 810             if (gmx_mm_any_lt(rsq02,rcutoff2))
 811             {
 812
 813             /* REACTION-FIELD ELECTROSTATICS */
 814             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_add_ps(rinv02,_mm_mul_ps(krf,rsq02)),crf));
 815             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
 816
 817             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 818
 819             /* Update potential sum for this i atom from the interaction with this j atom. */
 820             velec            = _mm_and_ps(velec,cutoff_mask);
 821             velec            = _mm_andnot_ps(dummy_mask,velec);
 822             velecsum         = _mm_add_ps(velecsum,velec);
 823
 824             fscal            = felec;
 825
 826             fscal            = _mm_and_ps(fscal,cutoff_mask);
 827
 828             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 829
 830             /* Calculate temporary vectorial force */
 831             tx               = _mm_mul_ps(fscal,dx02);
 832             ty               = _mm_mul_ps(fscal,dy02);
 833             tz               = _mm_mul_ps(fscal,dz02);
 834
 835             /* Update vectorial force */
 836             fix0             = _mm_add_ps(fix0,tx);
 837             fiy0             = _mm_add_ps(fiy0,ty);
 838             fiz0             = _mm_add_ps(fiz0,tz);
 839
 840             fjx2             = _mm_add_ps(fjx2,tx);
 841             fjy2             = _mm_add_ps(fjy2,ty);
 842             fjz2             = _mm_add_ps(fjz2,tz);
 843
 844             }
 845
 846             /**************************
 847              * CALCULATE INTERACTIONS *
 848              **************************/
 849
 850             if (gmx_mm_any_lt(rsq10,rcutoff2))
 851             {
 852
 853             /* REACTION-FIELD ELECTROSTATICS */
 854             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_add_ps(rinv10,_mm_mul_ps(krf,rsq10)),crf));
 855             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
 856
 857             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 858
 859             /* Update potential sum for this i atom from the interaction with this j atom. */
 860             velec            = _mm_and_ps(velec,cutoff_mask);
 861             velec            = _mm_andnot_ps(dummy_mask,velec);
 862             velecsum         = _mm_add_ps(velecsum,velec);
 863
 864             fscal            = felec;
 865
 866             fscal            = _mm_and_ps(fscal,cutoff_mask);
 867
 868             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 869
 870             /* Calculate temporary vectorial force */
 871             tx               = _mm_mul_ps(fscal,dx10);
 872             ty               = _mm_mul_ps(fscal,dy10);
 873             tz               = _mm_mul_ps(fscal,dz10);
 874
 875             /* Update vectorial force */
 876             fix1             = _mm_add_ps(fix1,tx);
 877             fiy1             = _mm_add_ps(fiy1,ty);
 878             fiz1             = _mm_add_ps(fiz1,tz);
 879
 880             fjx0             = _mm_add_ps(fjx0,tx);
 881             fjy0             = _mm_add_ps(fjy0,ty);
 882             fjz0             = _mm_add_ps(fjz0,tz);
 883
 884             }
 885
 886             /**************************
 887              * CALCULATE INTERACTIONS *
 888              **************************/
 889
 890             if (gmx_mm_any_lt(rsq11,rcutoff2))
 891             {
 892
 893             /* REACTION-FIELD ELECTROSTATICS */
 894             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 895             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 896
 897             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 898
 899             /* Update potential sum for this i atom from the interaction with this j atom. */
 900             velec            = _mm_and_ps(velec,cutoff_mask);
 901             velec            = _mm_andnot_ps(dummy_mask,velec);
 902             velecsum         = _mm_add_ps(velecsum,velec);
 903
 904             fscal            = felec;
 905
 906             fscal            = _mm_and_ps(fscal,cutoff_mask);
 907
 908             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 909
 910             /* Calculate temporary vectorial force */
 911             tx               = _mm_mul_ps(fscal,dx11);
 912             ty               = _mm_mul_ps(fscal,dy11);
 913             tz               = _mm_mul_ps(fscal,dz11);
 914
 915             /* Update vectorial force */
 916             fix1             = _mm_add_ps(fix1,tx);
 917             fiy1             = _mm_add_ps(fiy1,ty);
 918             fiz1             = _mm_add_ps(fiz1,tz);
 919
 920             fjx1             = _mm_add_ps(fjx1,tx);
 921             fjy1             = _mm_add_ps(fjy1,ty);
 922             fjz1             = _mm_add_ps(fjz1,tz);
 923
 924             }
 925
 926             /**************************
 927              * CALCULATE INTERACTIONS *
 928              **************************/
 929
 930             if (gmx_mm_any_lt(rsq12,rcutoff2))
 931             {
 932
 933             /* REACTION-FIELD ELECTROSTATICS */
 934             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 935             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 936
 937             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 938
 939             /* Update potential sum for this i atom from the interaction with this j atom. */
 940             velec            = _mm_and_ps(velec,cutoff_mask);
 941             velec            = _mm_andnot_ps(dummy_mask,velec);
 942             velecsum         = _mm_add_ps(velecsum,velec);
 943
 944             fscal            = felec;
 945
 946             fscal            = _mm_and_ps(fscal,cutoff_mask);
 947
 948             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 949
 950             /* Calculate temporary vectorial force */
 951             tx               = _mm_mul_ps(fscal,dx12);
 952             ty               = _mm_mul_ps(fscal,dy12);
 953             tz               = _mm_mul_ps(fscal,dz12);
 954
 955             /* Update vectorial force */
 956             fix1             = _mm_add_ps(fix1,tx);
 957             fiy1             = _mm_add_ps(fiy1,ty);
 958             fiz1             = _mm_add_ps(fiz1,tz);
 959
 960             fjx2             = _mm_add_ps(fjx2,tx);
 961             fjy2             = _mm_add_ps(fjy2,ty);
 962             fjz2             = _mm_add_ps(fjz2,tz);
 963
 964             }
 965
 966             /**************************
 967              * CALCULATE INTERACTIONS *
 968              **************************/
 969
 970             if (gmx_mm_any_lt(rsq20,rcutoff2))
 971             {
 972
 973             /* REACTION-FIELD ELECTROSTATICS */
 974             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_add_ps(rinv20,_mm_mul_ps(krf,rsq20)),crf));
 975             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
 976
 977             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 978
 979             /* Update potential sum for this i atom from the interaction with this j atom. */
 980             velec            = _mm_and_ps(velec,cutoff_mask);
 981             velec            = _mm_andnot_ps(dummy_mask,velec);
 982             velecsum         = _mm_add_ps(velecsum,velec);
 983
 984             fscal            = felec;
 985
 986             fscal            = _mm_and_ps(fscal,cutoff_mask);
 987
 988             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 989
 990             /* Calculate temporary vectorial force */
 991             tx               = _mm_mul_ps(fscal,dx20);
 992             ty               = _mm_mul_ps(fscal,dy20);
 993             tz               = _mm_mul_ps(fscal,dz20);
 994
 995             /* Update vectorial force */
 996             fix2             = _mm_add_ps(fix2,tx);
 997             fiy2             = _mm_add_ps(fiy2,ty);
 998             fiz2             = _mm_add_ps(fiz2,tz);
 999
1000             fjx0             = _mm_add_ps(fjx0,tx);
1001             fjy0             = _mm_add_ps(fjy0,ty);
1002             fjz0             = _mm_add_ps(fjz0,tz);
1003
1004             }
1005
1006             /**************************
1007              * CALCULATE INTERACTIONS *
1008              **************************/
1009
1010             if (gmx_mm_any_lt(rsq21,rcutoff2))
1011             {
1012
1013             /* REACTION-FIELD ELECTROSTATICS */
1014             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
1015             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1016
1017             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1018
1019             /* Update potential sum for this i atom from the interaction with this j atom. */
1020             velec            = _mm_and_ps(velec,cutoff_mask);
1021             velec            = _mm_andnot_ps(dummy_mask,velec);
1022             velecsum         = _mm_add_ps(velecsum,velec);
1023
1024             fscal            = felec;
1025
1026             fscal            = _mm_and_ps(fscal,cutoff_mask);
1027
1028             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1029
1030             /* Calculate temporary vectorial force */
1031             tx               = _mm_mul_ps(fscal,dx21);
1032             ty               = _mm_mul_ps(fscal,dy21);
1033             tz               = _mm_mul_ps(fscal,dz21);
1034
1035             /* Update vectorial force */
1036             fix2             = _mm_add_ps(fix2,tx);
1037             fiy2             = _mm_add_ps(fiy2,ty);
1038             fiz2             = _mm_add_ps(fiz2,tz);
1039
1040             fjx1             = _mm_add_ps(fjx1,tx);
1041             fjy1             = _mm_add_ps(fjy1,ty);
1042             fjz1             = _mm_add_ps(fjz1,tz);
1043
1044             }
1045
1046             /**************************
1047              * CALCULATE INTERACTIONS *
1048              **************************/
1049
1050             if (gmx_mm_any_lt(rsq22,rcutoff2))
1051             {
1052
1053             /* REACTION-FIELD ELECTROSTATICS */
1054             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
1055             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1056
1057             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1058
1059             /* Update potential sum for this i atom from the interaction with this j atom. */
1060             velec            = _mm_and_ps(velec,cutoff_mask);
1061             velec            = _mm_andnot_ps(dummy_mask,velec);
1062             velecsum         = _mm_add_ps(velecsum,velec);
1063
1064             fscal            = felec;
1065
1066             fscal            = _mm_and_ps(fscal,cutoff_mask);
1067
1068             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1069
1070             /* Calculate temporary vectorial force */
1071             tx               = _mm_mul_ps(fscal,dx22);
1072             ty               = _mm_mul_ps(fscal,dy22);
1073             tz               = _mm_mul_ps(fscal,dz22);
1074
1075             /* Update vectorial force */
1076             fix2             = _mm_add_ps(fix2,tx);
1077             fiy2             = _mm_add_ps(fiy2,ty);
1078             fiz2             = _mm_add_ps(fiz2,tz);
1079
1080             fjx2             = _mm_add_ps(fjx2,tx);
1081             fjy2             = _mm_add_ps(fjy2,ty);
1082             fjz2             = _mm_add_ps(fjz2,tz);
1083
1084             }
1085
1086             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1087             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1088             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1089             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1090
1091             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1092                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1093
1094             /* Inner loop uses 324 flops */
1095         }
1096
1097         /* End of innermost loop */
1098
1099         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1100                                               f+i_coord_offset,fshift+i_shift_offset);
1101
1102         ggid                        = gid[iidx];
1103         /* Update potential energies */
1104         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1105
1106         /* Increment number of inner iterations */
1107         inneriter                  += j_index_end - j_index_start;
1108
1109         /* Outer loop uses 19 flops */
1110     }
1111
1112     /* Increment number of outer iterations */
1113     outeriter        += nri;
1114
1115     /* Update outer/inner flops */
1116
1117     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_VF,outeriter*19 + inneriter*324);
1118 }
1119 /*
1120  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwNone_GeomW3W3_F_sse2_single
1121  * Electrostatics interaction: ReactionField
1122  * VdW interaction:            None
1123  * Geometry:                   Water3-Water3
1124  * Calculate force/pot:        Force
1125  */
1126 void
1127 nb_kernel_ElecRFCut_VdwNone_GeomW3W3_F_sse2_single
1128                     (t_nblist                    * gmx_restrict       nlist,
1129                      rvec                        * gmx_restrict          xx,
1130                      rvec                        * gmx_restrict          ff,
1131                      struct t_forcerec           * gmx_restrict          fr,
1132                      t_mdatoms                   * gmx_restrict     mdatoms,
1133                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1134                      t_nrnb                      * gmx_restrict        nrnb)
1135 {
1136     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1137      * just 0 for non-waters.
1138      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
1139      * jnr indices corresponding to data put in the four positions in the SIMD register.
1140      */
1141     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1142     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1143     int              jnrA,jnrB,jnrC,jnrD;
1144     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1145     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1146     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1147     real             rcutoff_scalar;
1148     real             *shiftvec,*fshift,*x,*f;
1149     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1150     real             scratch[4*DIM];
1151     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1152     int              vdwioffset0;
1153     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1154     int              vdwioffset1;
1155     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1156     int              vdwioffset2;
1157     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1158     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1159     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1160     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1161     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1162     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1163     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1164     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1165     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1166     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1167     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1168     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1169     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1170     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1171     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1172     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1173     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1174     real             *charge;
1175     __m128           dummy_mask,cutoff_mask;
1176     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1177     __m128           one     = _mm_set1_ps(1.0);
1178     __m128           two     = _mm_set1_ps(2.0);
1179     x                = xx[0];
1180     f                = ff[0];
1181
1182     nri              = nlist->nri;
1183     iinr             = nlist->iinr;
1184     jindex           = nlist->jindex;
1185     jjnr             = nlist->jjnr;
1186     shiftidx         = nlist->shift;
1187     gid              = nlist->gid;
1188     shiftvec         = fr->shift_vec[0];
1189     fshift           = fr->fshift[0];
1190     facel            = _mm_set1_ps(fr->ic->epsfac);
1191     charge           = mdatoms->chargeA;
1192     krf              = _mm_set1_ps(fr->ic->k_rf);
1193     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1194     crf              = _mm_set1_ps(fr->ic->c_rf);
1195
1196     /* Setup water-specific parameters */
1197     inr              = nlist->iinr[0];
1198     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+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
1202     jq0              = _mm_set1_ps(charge[inr+0]);
1203     jq1              = _mm_set1_ps(charge[inr+1]);
1204     jq2              = _mm_set1_ps(charge[inr+2]);
1205     qq00             = _mm_mul_ps(iq0,jq0);
1206     qq01             = _mm_mul_ps(iq0,jq1);
1207     qq02             = _mm_mul_ps(iq0,jq2);
1208     qq10             = _mm_mul_ps(iq1,jq0);
1209     qq11             = _mm_mul_ps(iq1,jq1);
1210     qq12             = _mm_mul_ps(iq1,jq2);
1211     qq20             = _mm_mul_ps(iq2,jq0);
1212     qq21             = _mm_mul_ps(iq2,jq1);
1213     qq22             = _mm_mul_ps(iq2,jq2);
1214
1215     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1216     rcutoff_scalar   = fr->ic->rcoulomb;
1217     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1218     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1219
1220     /* Avoid stupid compiler warnings */
1221     jnrA = jnrB = jnrC = jnrD = 0;
1222     j_coord_offsetA = 0;
1223     j_coord_offsetB = 0;
1224     j_coord_offsetC = 0;
1225     j_coord_offsetD = 0;
1226
1227     outeriter        = 0;
1228     inneriter        = 0;
1229
1230     for(iidx=0;iidx<4*DIM;iidx++)
1231     {
1232         scratch[iidx] = 0.0;
1233     }
1234
1235     /* Start outer loop over neighborlists */
1236     for(iidx=0; iidx<nri; iidx++)
1237     {
1238         /* Load shift vector for this list */
1239         i_shift_offset   = DIM*shiftidx[iidx];
1240
1241         /* Load limits for loop over neighbors */
1242         j_index_start    = jindex[iidx];
1243         j_index_end      = jindex[iidx+1];
1244
1245         /* Get outer coordinate index */
1246         inr              = iinr[iidx];
1247         i_coord_offset   = DIM*inr;
1248
1249         /* Load i particle coords and add shift vector */
1250         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1251                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1252
1253         fix0             = _mm_setzero_ps();
1254         fiy0             = _mm_setzero_ps();
1255         fiz0             = _mm_setzero_ps();
1256         fix1             = _mm_setzero_ps();
1257         fiy1             = _mm_setzero_ps();
1258         fiz1             = _mm_setzero_ps();
1259         fix2             = _mm_setzero_ps();
1260         fiy2             = _mm_setzero_ps();
1261         fiz2             = _mm_setzero_ps();
1262
1263         /* Start inner kernel loop */
1264         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1265         {
1266
1267             /* Get j neighbor index, and coordinate index */
1268             jnrA             = jjnr[jidx];
1269             jnrB             = jjnr[jidx+1];
1270             jnrC             = jjnr[jidx+2];
1271             jnrD             = jjnr[jidx+3];
1272             j_coord_offsetA  = DIM*jnrA;
1273             j_coord_offsetB  = DIM*jnrB;
1274             j_coord_offsetC  = DIM*jnrC;
1275             j_coord_offsetD  = DIM*jnrD;
1276
1277             /* load j atom coordinates */
1278             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1279                                               x+j_coord_offsetC,x+j_coord_offsetD,
1280                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1281
1282             /* Calculate displacement vector */
1283             dx00             = _mm_sub_ps(ix0,jx0);
1284             dy00             = _mm_sub_ps(iy0,jy0);
1285             dz00             = _mm_sub_ps(iz0,jz0);
1286             dx01             = _mm_sub_ps(ix0,jx1);
1287             dy01             = _mm_sub_ps(iy0,jy1);
1288             dz01             = _mm_sub_ps(iz0,jz1);
1289             dx02             = _mm_sub_ps(ix0,jx2);
1290             dy02             = _mm_sub_ps(iy0,jy2);
1291             dz02             = _mm_sub_ps(iz0,jz2);
1292             dx10             = _mm_sub_ps(ix1,jx0);
1293             dy10             = _mm_sub_ps(iy1,jy0);
1294             dz10             = _mm_sub_ps(iz1,jz0);
1295             dx11             = _mm_sub_ps(ix1,jx1);
1296             dy11             = _mm_sub_ps(iy1,jy1);
1297             dz11             = _mm_sub_ps(iz1,jz1);
1298             dx12             = _mm_sub_ps(ix1,jx2);
1299             dy12             = _mm_sub_ps(iy1,jy2);
1300             dz12             = _mm_sub_ps(iz1,jz2);
1301             dx20             = _mm_sub_ps(ix2,jx0);
1302             dy20             = _mm_sub_ps(iy2,jy0);
1303             dz20             = _mm_sub_ps(iz2,jz0);
1304             dx21             = _mm_sub_ps(ix2,jx1);
1305             dy21             = _mm_sub_ps(iy2,jy1);
1306             dz21             = _mm_sub_ps(iz2,jz1);
1307             dx22             = _mm_sub_ps(ix2,jx2);
1308             dy22             = _mm_sub_ps(iy2,jy2);
1309             dz22             = _mm_sub_ps(iz2,jz2);
1310
1311             /* Calculate squared distance and things based on it */
1312             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1313             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1314             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1315             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1316             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1317             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1318             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1319             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1320             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1321
1322             rinv00           = sse2_invsqrt_f(rsq00);
1323             rinv01           = sse2_invsqrt_f(rsq01);
1324             rinv02           = sse2_invsqrt_f(rsq02);
1325             rinv10           = sse2_invsqrt_f(rsq10);
1326             rinv11           = sse2_invsqrt_f(rsq11);
1327             rinv12           = sse2_invsqrt_f(rsq12);
1328             rinv20           = sse2_invsqrt_f(rsq20);
1329             rinv21           = sse2_invsqrt_f(rsq21);
1330             rinv22           = sse2_invsqrt_f(rsq22);
1331
1332             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1333             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1334             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1335             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1336             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1337             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1338             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1339             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1340             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1341
1342             fjx0             = _mm_setzero_ps();
1343             fjy0             = _mm_setzero_ps();
1344             fjz0             = _mm_setzero_ps();
1345             fjx1             = _mm_setzero_ps();
1346             fjy1             = _mm_setzero_ps();
1347             fjz1             = _mm_setzero_ps();
1348             fjx2             = _mm_setzero_ps();
1349             fjy2             = _mm_setzero_ps();
1350             fjz2             = _mm_setzero_ps();
1351
1352             /**************************
1353              * CALCULATE INTERACTIONS *
1354              **************************/
1355
1356             if (gmx_mm_any_lt(rsq00,rcutoff2))
1357             {
1358
1359             /* REACTION-FIELD ELECTROSTATICS */
1360             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
1361
1362             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1363
1364             fscal            = felec;
1365
1366             fscal            = _mm_and_ps(fscal,cutoff_mask);
1367
1368             /* Calculate temporary vectorial force */
1369             tx               = _mm_mul_ps(fscal,dx00);
1370             ty               = _mm_mul_ps(fscal,dy00);
1371             tz               = _mm_mul_ps(fscal,dz00);
1372
1373             /* Update vectorial force */
1374             fix0             = _mm_add_ps(fix0,tx);
1375             fiy0             = _mm_add_ps(fiy0,ty);
1376             fiz0             = _mm_add_ps(fiz0,tz);
1377
1378             fjx0             = _mm_add_ps(fjx0,tx);
1379             fjy0             = _mm_add_ps(fjy0,ty);
1380             fjz0             = _mm_add_ps(fjz0,tz);
1381
1382             }
1383
1384             /**************************
1385              * CALCULATE INTERACTIONS *
1386              **************************/
1387
1388             if (gmx_mm_any_lt(rsq01,rcutoff2))
1389             {
1390
1391             /* REACTION-FIELD ELECTROSTATICS */
1392             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
1393
1394             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1395
1396             fscal            = felec;
1397
1398             fscal            = _mm_and_ps(fscal,cutoff_mask);
1399
1400             /* Calculate temporary vectorial force */
1401             tx               = _mm_mul_ps(fscal,dx01);
1402             ty               = _mm_mul_ps(fscal,dy01);
1403             tz               = _mm_mul_ps(fscal,dz01);
1404
1405             /* Update vectorial force */
1406             fix0             = _mm_add_ps(fix0,tx);
1407             fiy0             = _mm_add_ps(fiy0,ty);
1408             fiz0             = _mm_add_ps(fiz0,tz);
1409
1410             fjx1             = _mm_add_ps(fjx1,tx);
1411             fjy1             = _mm_add_ps(fjy1,ty);
1412             fjz1             = _mm_add_ps(fjz1,tz);
1413
1414             }
1415
1416             /**************************
1417              * CALCULATE INTERACTIONS *
1418              **************************/
1419
1420             if (gmx_mm_any_lt(rsq02,rcutoff2))
1421             {
1422
1423             /* REACTION-FIELD ELECTROSTATICS */
1424             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
1425
1426             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
1427
1428             fscal            = felec;
1429
1430             fscal            = _mm_and_ps(fscal,cutoff_mask);
1431
1432             /* Calculate temporary vectorial force */
1433             tx               = _mm_mul_ps(fscal,dx02);
1434             ty               = _mm_mul_ps(fscal,dy02);
1435             tz               = _mm_mul_ps(fscal,dz02);
1436
1437             /* Update vectorial force */
1438             fix0             = _mm_add_ps(fix0,tx);
1439             fiy0             = _mm_add_ps(fiy0,ty);
1440             fiz0             = _mm_add_ps(fiz0,tz);
1441
1442             fjx2             = _mm_add_ps(fjx2,tx);
1443             fjy2             = _mm_add_ps(fjy2,ty);
1444             fjz2             = _mm_add_ps(fjz2,tz);
1445
1446             }
1447
1448             /**************************
1449              * CALCULATE INTERACTIONS *
1450              **************************/
1451
1452             if (gmx_mm_any_lt(rsq10,rcutoff2))
1453             {
1454
1455             /* REACTION-FIELD ELECTROSTATICS */
1456             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
1457
1458             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1459
1460             fscal            = felec;
1461
1462             fscal            = _mm_and_ps(fscal,cutoff_mask);
1463
1464             /* Calculate temporary vectorial force */
1465             tx               = _mm_mul_ps(fscal,dx10);
1466             ty               = _mm_mul_ps(fscal,dy10);
1467             tz               = _mm_mul_ps(fscal,dz10);
1468
1469             /* Update vectorial force */
1470             fix1             = _mm_add_ps(fix1,tx);
1471             fiy1             = _mm_add_ps(fiy1,ty);
1472             fiz1             = _mm_add_ps(fiz1,tz);
1473
1474             fjx0             = _mm_add_ps(fjx0,tx);
1475             fjy0             = _mm_add_ps(fjy0,ty);
1476             fjz0             = _mm_add_ps(fjz0,tz);
1477
1478             }
1479
1480             /**************************
1481              * CALCULATE INTERACTIONS *
1482              **************************/
1483
1484             if (gmx_mm_any_lt(rsq11,rcutoff2))
1485             {
1486
1487             /* REACTION-FIELD ELECTROSTATICS */
1488             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1489
1490             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1491
1492             fscal            = felec;
1493
1494             fscal            = _mm_and_ps(fscal,cutoff_mask);
1495
1496             /* Calculate temporary vectorial force */
1497             tx               = _mm_mul_ps(fscal,dx11);
1498             ty               = _mm_mul_ps(fscal,dy11);
1499             tz               = _mm_mul_ps(fscal,dz11);
1500
1501             /* Update vectorial force */
1502             fix1             = _mm_add_ps(fix1,tx);
1503             fiy1             = _mm_add_ps(fiy1,ty);
1504             fiz1             = _mm_add_ps(fiz1,tz);
1505
1506             fjx1             = _mm_add_ps(fjx1,tx);
1507             fjy1             = _mm_add_ps(fjy1,ty);
1508             fjz1             = _mm_add_ps(fjz1,tz);
1509
1510             }
1511
1512             /**************************
1513              * CALCULATE INTERACTIONS *
1514              **************************/
1515
1516             if (gmx_mm_any_lt(rsq12,rcutoff2))
1517             {
1518
1519             /* REACTION-FIELD ELECTROSTATICS */
1520             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1521
1522             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1523
1524             fscal            = felec;
1525
1526             fscal            = _mm_and_ps(fscal,cutoff_mask);
1527
1528             /* Calculate temporary vectorial force */
1529             tx               = _mm_mul_ps(fscal,dx12);
1530             ty               = _mm_mul_ps(fscal,dy12);
1531             tz               = _mm_mul_ps(fscal,dz12);
1532
1533             /* Update vectorial force */
1534             fix1             = _mm_add_ps(fix1,tx);
1535             fiy1             = _mm_add_ps(fiy1,ty);
1536             fiz1             = _mm_add_ps(fiz1,tz);
1537
1538             fjx2             = _mm_add_ps(fjx2,tx);
1539             fjy2             = _mm_add_ps(fjy2,ty);
1540             fjz2             = _mm_add_ps(fjz2,tz);
1541
1542             }
1543
1544             /**************************
1545              * CALCULATE INTERACTIONS *
1546              **************************/
1547
1548             if (gmx_mm_any_lt(rsq20,rcutoff2))
1549             {
1550
1551             /* REACTION-FIELD ELECTROSTATICS */
1552             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
1553
1554             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1555
1556             fscal            = felec;
1557
1558             fscal            = _mm_and_ps(fscal,cutoff_mask);
1559
1560             /* Calculate temporary vectorial force */
1561             tx               = _mm_mul_ps(fscal,dx20);
1562             ty               = _mm_mul_ps(fscal,dy20);
1563             tz               = _mm_mul_ps(fscal,dz20);
1564
1565             /* Update vectorial force */
1566             fix2             = _mm_add_ps(fix2,tx);
1567             fiy2             = _mm_add_ps(fiy2,ty);
1568             fiz2             = _mm_add_ps(fiz2,tz);
1569
1570             fjx0             = _mm_add_ps(fjx0,tx);
1571             fjy0             = _mm_add_ps(fjy0,ty);
1572             fjz0             = _mm_add_ps(fjz0,tz);
1573
1574             }
1575
1576             /**************************
1577              * CALCULATE INTERACTIONS *
1578              **************************/
1579
1580             if (gmx_mm_any_lt(rsq21,rcutoff2))
1581             {
1582
1583             /* REACTION-FIELD ELECTROSTATICS */
1584             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1585
1586             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1587
1588             fscal            = felec;
1589
1590             fscal            = _mm_and_ps(fscal,cutoff_mask);
1591
1592             /* Calculate temporary vectorial force */
1593             tx               = _mm_mul_ps(fscal,dx21);
1594             ty               = _mm_mul_ps(fscal,dy21);
1595             tz               = _mm_mul_ps(fscal,dz21);
1596
1597             /* Update vectorial force */
1598             fix2             = _mm_add_ps(fix2,tx);
1599             fiy2             = _mm_add_ps(fiy2,ty);
1600             fiz2             = _mm_add_ps(fiz2,tz);
1601
1602             fjx1             = _mm_add_ps(fjx1,tx);
1603             fjy1             = _mm_add_ps(fjy1,ty);
1604             fjz1             = _mm_add_ps(fjz1,tz);
1605
1606             }
1607
1608             /**************************
1609              * CALCULATE INTERACTIONS *
1610              **************************/
1611
1612             if (gmx_mm_any_lt(rsq22,rcutoff2))
1613             {
1614
1615             /* REACTION-FIELD ELECTROSTATICS */
1616             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1617
1618             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1619
1620             fscal            = felec;
1621
1622             fscal            = _mm_and_ps(fscal,cutoff_mask);
1623
1624             /* Calculate temporary vectorial force */
1625             tx               = _mm_mul_ps(fscal,dx22);
1626             ty               = _mm_mul_ps(fscal,dy22);
1627             tz               = _mm_mul_ps(fscal,dz22);
1628
1629             /* Update vectorial force */
1630             fix2             = _mm_add_ps(fix2,tx);
1631             fiy2             = _mm_add_ps(fiy2,ty);
1632             fiz2             = _mm_add_ps(fiz2,tz);
1633
1634             fjx2             = _mm_add_ps(fjx2,tx);
1635             fjy2             = _mm_add_ps(fjy2,ty);
1636             fjz2             = _mm_add_ps(fjz2,tz);
1637
1638             }
1639
1640             fjptrA             = f+j_coord_offsetA;
1641             fjptrB             = f+j_coord_offsetB;
1642             fjptrC             = f+j_coord_offsetC;
1643             fjptrD             = f+j_coord_offsetD;
1644
1645             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1646                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1647
1648             /* Inner loop uses 270 flops */
1649         }
1650
1651         if(jidx<j_index_end)
1652         {
1653
1654             /* Get j neighbor index, and coordinate index */
1655             jnrlistA         = jjnr[jidx];
1656             jnrlistB         = jjnr[jidx+1];
1657             jnrlistC         = jjnr[jidx+2];
1658             jnrlistD         = jjnr[jidx+3];
1659             /* Sign of each element will be negative for non-real atoms.
1660              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1661              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1662              */
1663             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1664             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1665             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1666             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1667             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1668             j_coord_offsetA  = DIM*jnrA;
1669             j_coord_offsetB  = DIM*jnrB;
1670             j_coord_offsetC  = DIM*jnrC;
1671             j_coord_offsetD  = DIM*jnrD;
1672
1673             /* load j atom coordinates */
1674             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1675                                               x+j_coord_offsetC,x+j_coord_offsetD,
1676                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1677
1678             /* Calculate displacement vector */
1679             dx00             = _mm_sub_ps(ix0,jx0);
1680             dy00             = _mm_sub_ps(iy0,jy0);
1681             dz00             = _mm_sub_ps(iz0,jz0);
1682             dx01             = _mm_sub_ps(ix0,jx1);
1683             dy01             = _mm_sub_ps(iy0,jy1);
1684             dz01             = _mm_sub_ps(iz0,jz1);
1685             dx02             = _mm_sub_ps(ix0,jx2);
1686             dy02             = _mm_sub_ps(iy0,jy2);
1687             dz02             = _mm_sub_ps(iz0,jz2);
1688             dx10             = _mm_sub_ps(ix1,jx0);
1689             dy10             = _mm_sub_ps(iy1,jy0);
1690             dz10             = _mm_sub_ps(iz1,jz0);
1691             dx11             = _mm_sub_ps(ix1,jx1);
1692             dy11             = _mm_sub_ps(iy1,jy1);
1693             dz11             = _mm_sub_ps(iz1,jz1);
1694             dx12             = _mm_sub_ps(ix1,jx2);
1695             dy12             = _mm_sub_ps(iy1,jy2);
1696             dz12             = _mm_sub_ps(iz1,jz2);
1697             dx20             = _mm_sub_ps(ix2,jx0);
1698             dy20             = _mm_sub_ps(iy2,jy0);
1699             dz20             = _mm_sub_ps(iz2,jz0);
1700             dx21             = _mm_sub_ps(ix2,jx1);
1701             dy21             = _mm_sub_ps(iy2,jy1);
1702             dz21             = _mm_sub_ps(iz2,jz1);
1703             dx22             = _mm_sub_ps(ix2,jx2);
1704             dy22             = _mm_sub_ps(iy2,jy2);
1705             dz22             = _mm_sub_ps(iz2,jz2);
1706
1707             /* Calculate squared distance and things based on it */
1708             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1709             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1710             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1711             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1712             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1713             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1714             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1715             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1716             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1717
1718             rinv00           = sse2_invsqrt_f(rsq00);
1719             rinv01           = sse2_invsqrt_f(rsq01);
1720             rinv02           = sse2_invsqrt_f(rsq02);
1721             rinv10           = sse2_invsqrt_f(rsq10);
1722             rinv11           = sse2_invsqrt_f(rsq11);
1723             rinv12           = sse2_invsqrt_f(rsq12);
1724             rinv20           = sse2_invsqrt_f(rsq20);
1725             rinv21           = sse2_invsqrt_f(rsq21);
1726             rinv22           = sse2_invsqrt_f(rsq22);
1727
1728             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1729             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1730             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1731             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1732             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1733             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1734             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1735             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1736             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1737
1738             fjx0             = _mm_setzero_ps();
1739             fjy0             = _mm_setzero_ps();
1740             fjz0             = _mm_setzero_ps();
1741             fjx1             = _mm_setzero_ps();
1742             fjy1             = _mm_setzero_ps();
1743             fjz1             = _mm_setzero_ps();
1744             fjx2             = _mm_setzero_ps();
1745             fjy2             = _mm_setzero_ps();
1746             fjz2             = _mm_setzero_ps();
1747
1748             /**************************
1749              * CALCULATE INTERACTIONS *
1750              **************************/
1751
1752             if (gmx_mm_any_lt(rsq00,rcutoff2))
1753             {
1754
1755             /* REACTION-FIELD ELECTROSTATICS */
1756             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
1757
1758             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1759
1760             fscal            = felec;
1761
1762             fscal            = _mm_and_ps(fscal,cutoff_mask);
1763
1764             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1765
1766             /* Calculate temporary vectorial force */
1767             tx               = _mm_mul_ps(fscal,dx00);
1768             ty               = _mm_mul_ps(fscal,dy00);
1769             tz               = _mm_mul_ps(fscal,dz00);
1770
1771             /* Update vectorial force */
1772             fix0             = _mm_add_ps(fix0,tx);
1773             fiy0             = _mm_add_ps(fiy0,ty);
1774             fiz0             = _mm_add_ps(fiz0,tz);
1775
1776             fjx0             = _mm_add_ps(fjx0,tx);
1777             fjy0             = _mm_add_ps(fjy0,ty);
1778             fjz0             = _mm_add_ps(fjz0,tz);
1779
1780             }
1781
1782             /**************************
1783              * CALCULATE INTERACTIONS *
1784              **************************/
1785
1786             if (gmx_mm_any_lt(rsq01,rcutoff2))
1787             {
1788
1789             /* REACTION-FIELD ELECTROSTATICS */
1790             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
1791
1792             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1793
1794             fscal            = felec;
1795
1796             fscal            = _mm_and_ps(fscal,cutoff_mask);
1797
1798             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1799
1800             /* Calculate temporary vectorial force */
1801             tx               = _mm_mul_ps(fscal,dx01);
1802             ty               = _mm_mul_ps(fscal,dy01);
1803             tz               = _mm_mul_ps(fscal,dz01);
1804
1805             /* Update vectorial force */
1806             fix0             = _mm_add_ps(fix0,tx);
1807             fiy0             = _mm_add_ps(fiy0,ty);
1808             fiz0             = _mm_add_ps(fiz0,tz);
1809
1810             fjx1             = _mm_add_ps(fjx1,tx);
1811             fjy1             = _mm_add_ps(fjy1,ty);
1812             fjz1             = _mm_add_ps(fjz1,tz);
1813
1814             }
1815
1816             /**************************
1817              * CALCULATE INTERACTIONS *
1818              **************************/
1819
1820             if (gmx_mm_any_lt(rsq02,rcutoff2))
1821             {
1822
1823             /* REACTION-FIELD ELECTROSTATICS */
1824             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
1825
1826             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
1827
1828             fscal            = felec;
1829
1830             fscal            = _mm_and_ps(fscal,cutoff_mask);
1831
1832             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1833
1834             /* Calculate temporary vectorial force */
1835             tx               = _mm_mul_ps(fscal,dx02);
1836             ty               = _mm_mul_ps(fscal,dy02);
1837             tz               = _mm_mul_ps(fscal,dz02);
1838
1839             /* Update vectorial force */
1840             fix0             = _mm_add_ps(fix0,tx);
1841             fiy0             = _mm_add_ps(fiy0,ty);
1842             fiz0             = _mm_add_ps(fiz0,tz);
1843
1844             fjx2             = _mm_add_ps(fjx2,tx);
1845             fjy2             = _mm_add_ps(fjy2,ty);
1846             fjz2             = _mm_add_ps(fjz2,tz);
1847
1848             }
1849
1850             /**************************
1851              * CALCULATE INTERACTIONS *
1852              **************************/
1853
1854             if (gmx_mm_any_lt(rsq10,rcutoff2))
1855             {
1856
1857             /* REACTION-FIELD ELECTROSTATICS */
1858             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
1859
1860             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1861
1862             fscal            = felec;
1863
1864             fscal            = _mm_and_ps(fscal,cutoff_mask);
1865
1866             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1867
1868             /* Calculate temporary vectorial force */
1869             tx               = _mm_mul_ps(fscal,dx10);
1870             ty               = _mm_mul_ps(fscal,dy10);
1871             tz               = _mm_mul_ps(fscal,dz10);
1872
1873             /* Update vectorial force */
1874             fix1             = _mm_add_ps(fix1,tx);
1875             fiy1             = _mm_add_ps(fiy1,ty);
1876             fiz1             = _mm_add_ps(fiz1,tz);
1877
1878             fjx0             = _mm_add_ps(fjx0,tx);
1879             fjy0             = _mm_add_ps(fjy0,ty);
1880             fjz0             = _mm_add_ps(fjz0,tz);
1881
1882             }
1883
1884             /**************************
1885              * CALCULATE INTERACTIONS *
1886              **************************/
1887
1888             if (gmx_mm_any_lt(rsq11,rcutoff2))
1889             {
1890
1891             /* REACTION-FIELD ELECTROSTATICS */
1892             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1893
1894             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1895
1896             fscal            = felec;
1897
1898             fscal            = _mm_and_ps(fscal,cutoff_mask);
1899
1900             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1901
1902             /* Calculate temporary vectorial force */
1903             tx               = _mm_mul_ps(fscal,dx11);
1904             ty               = _mm_mul_ps(fscal,dy11);
1905             tz               = _mm_mul_ps(fscal,dz11);
1906
1907             /* Update vectorial force */
1908             fix1             = _mm_add_ps(fix1,tx);
1909             fiy1             = _mm_add_ps(fiy1,ty);
1910             fiz1             = _mm_add_ps(fiz1,tz);
1911
1912             fjx1             = _mm_add_ps(fjx1,tx);
1913             fjy1             = _mm_add_ps(fjy1,ty);
1914             fjz1             = _mm_add_ps(fjz1,tz);
1915
1916             }
1917
1918             /**************************
1919              * CALCULATE INTERACTIONS *
1920              **************************/
1921
1922             if (gmx_mm_any_lt(rsq12,rcutoff2))
1923             {
1924
1925             /* REACTION-FIELD ELECTROSTATICS */
1926             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1927
1928             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1929
1930             fscal            = felec;
1931
1932             fscal            = _mm_and_ps(fscal,cutoff_mask);
1933
1934             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1935
1936             /* Calculate temporary vectorial force */
1937             tx               = _mm_mul_ps(fscal,dx12);
1938             ty               = _mm_mul_ps(fscal,dy12);
1939             tz               = _mm_mul_ps(fscal,dz12);
1940
1941             /* Update vectorial force */
1942             fix1             = _mm_add_ps(fix1,tx);
1943             fiy1             = _mm_add_ps(fiy1,ty);
1944             fiz1             = _mm_add_ps(fiz1,tz);
1945
1946             fjx2             = _mm_add_ps(fjx2,tx);
1947             fjy2             = _mm_add_ps(fjy2,ty);
1948             fjz2             = _mm_add_ps(fjz2,tz);
1949
1950             }
1951
1952             /**************************
1953              * CALCULATE INTERACTIONS *
1954              **************************/
1955
1956             if (gmx_mm_any_lt(rsq20,rcutoff2))
1957             {
1958
1959             /* REACTION-FIELD ELECTROSTATICS */
1960             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
1961
1962             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1963
1964             fscal            = felec;
1965
1966             fscal            = _mm_and_ps(fscal,cutoff_mask);
1967
1968             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1969
1970             /* Calculate temporary vectorial force */
1971             tx               = _mm_mul_ps(fscal,dx20);
1972             ty               = _mm_mul_ps(fscal,dy20);
1973             tz               = _mm_mul_ps(fscal,dz20);
1974
1975             /* Update vectorial force */
1976             fix2             = _mm_add_ps(fix2,tx);
1977             fiy2             = _mm_add_ps(fiy2,ty);
1978             fiz2             = _mm_add_ps(fiz2,tz);
1979
1980             fjx0             = _mm_add_ps(fjx0,tx);
1981             fjy0             = _mm_add_ps(fjy0,ty);
1982             fjz0             = _mm_add_ps(fjz0,tz);
1983
1984             }
1985
1986             /**************************
1987              * CALCULATE INTERACTIONS *
1988              **************************/
1989
1990             if (gmx_mm_any_lt(rsq21,rcutoff2))
1991             {
1992
1993             /* REACTION-FIELD ELECTROSTATICS */
1994             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1995
1996             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1997
1998             fscal            = felec;
1999
2000             fscal            = _mm_and_ps(fscal,cutoff_mask);
2001
2002             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2003
2004             /* Calculate temporary vectorial force */
2005             tx               = _mm_mul_ps(fscal,dx21);
2006             ty               = _mm_mul_ps(fscal,dy21);
2007             tz               = _mm_mul_ps(fscal,dz21);
2008
2009             /* Update vectorial force */
2010             fix2             = _mm_add_ps(fix2,tx);
2011             fiy2             = _mm_add_ps(fiy2,ty);
2012             fiz2             = _mm_add_ps(fiz2,tz);
2013
2014             fjx1             = _mm_add_ps(fjx1,tx);
2015             fjy1             = _mm_add_ps(fjy1,ty);
2016             fjz1             = _mm_add_ps(fjz1,tz);
2017
2018             }
2019
2020             /**************************
2021              * CALCULATE INTERACTIONS *
2022              **************************/
2023
2024             if (gmx_mm_any_lt(rsq22,rcutoff2))
2025             {
2026
2027             /* REACTION-FIELD ELECTROSTATICS */
2028             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
2029
2030             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
2031
2032             fscal            = felec;
2033
2034             fscal            = _mm_and_ps(fscal,cutoff_mask);
2035
2036             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2037
2038             /* Calculate temporary vectorial force */
2039             tx               = _mm_mul_ps(fscal,dx22);
2040             ty               = _mm_mul_ps(fscal,dy22);
2041             tz               = _mm_mul_ps(fscal,dz22);
2042
2043             /* Update vectorial force */
2044             fix2             = _mm_add_ps(fix2,tx);
2045             fiy2             = _mm_add_ps(fiy2,ty);
2046             fiz2             = _mm_add_ps(fiz2,tz);
2047
2048             fjx2             = _mm_add_ps(fjx2,tx);
2049             fjy2             = _mm_add_ps(fjy2,ty);
2050             fjz2             = _mm_add_ps(fjz2,tz);
2051
2052             }
2053
2054             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2055             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2056             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2057             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2058
2059             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
2060                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2061
2062             /* Inner loop uses 270 flops */
2063         }
2064
2065         /* End of innermost loop */
2066
2067         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
2068                                               f+i_coord_offset,fshift+i_shift_offset);
2069
2070         /* Increment number of inner iterations */
2071         inneriter                  += j_index_end - j_index_start;
2072
2073         /* Outer loop uses 18 flops */
2074     }
2075
2076     /* Increment number of outer iterations */
2077     outeriter        += nri;
2078
2079     /* Update outer/inner flops */
2080
2081     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_F,outeriter*18 + inneriter*270);
2082 }