src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_single/nb_kernel_ElecRF_VdwNone_GeomW3W3_sse4_1_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS sse4_1_single kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_sse4_1_single.h"
  48 #include "kernelutil_x86_sse4_1_single.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwNone_GeomW3W3_VF_sse4_1_single
  52  * Electrostatics interaction: ReactionField
  53  * VdW interaction:            None
  54  * Geometry:                   Water3-Water3
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecRF_VdwNone_GeomW3W3_VF_sse4_1_single
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB,jnrC,jnrD;
  75     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  76     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  77     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  78     real             rcutoff_scalar;
  79     real             *shiftvec,*fshift,*x,*f;
  80     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  81     real             scratch[4*DIM];
  82     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  83     int              vdwioffset0;
  84     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  85     int              vdwioffset1;
  86     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  87     int              vdwioffset2;
  88     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  89     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  90     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  91     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  92     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  93     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  94     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  95     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  96     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  97     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  98     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  99     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 100     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 101     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 102     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 103     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 104     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 105     real             *charge;
 106     __m128           dummy_mask,cutoff_mask;
 107     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 108     __m128           one     = _mm_set1_ps(1.0);
 109     __m128           two     = _mm_set1_ps(2.0);
 110     x                = xx[0];
 111     f                = ff[0];
 112
 113     nri              = nlist->nri;
 114     iinr             = nlist->iinr;
 115     jindex           = nlist->jindex;
 116     jjnr             = nlist->jjnr;
 117     shiftidx         = nlist->shift;
 118     gid              = nlist->gid;
 119     shiftvec         = fr->shift_vec[0];
 120     fshift           = fr->fshift[0];
 121     facel            = _mm_set1_ps(fr->epsfac);
 122     charge           = mdatoms->chargeA;
 123     krf              = _mm_set1_ps(fr->ic->k_rf);
 124     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 125     crf              = _mm_set1_ps(fr->ic->c_rf);
 126
 127     /* Setup water-specific parameters */
 128     inr              = nlist->iinr[0];
 129     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 130     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 131     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 132
 133     jq0              = _mm_set1_ps(charge[inr+0]);
 134     jq1              = _mm_set1_ps(charge[inr+1]);
 135     jq2              = _mm_set1_ps(charge[inr+2]);
 136     qq00             = _mm_mul_ps(iq0,jq0);
 137     qq01             = _mm_mul_ps(iq0,jq1);
 138     qq02             = _mm_mul_ps(iq0,jq2);
 139     qq10             = _mm_mul_ps(iq1,jq0);
 140     qq11             = _mm_mul_ps(iq1,jq1);
 141     qq12             = _mm_mul_ps(iq1,jq2);
 142     qq20             = _mm_mul_ps(iq2,jq0);
 143     qq21             = _mm_mul_ps(iq2,jq1);
 144     qq22             = _mm_mul_ps(iq2,jq2);
 145
 146     /* Avoid stupid compiler warnings */
 147     jnrA = jnrB = jnrC = jnrD = 0;
 148     j_coord_offsetA = 0;
 149     j_coord_offsetB = 0;
 150     j_coord_offsetC = 0;
 151     j_coord_offsetD = 0;
 152
 153     outeriter        = 0;
 154     inneriter        = 0;
 155
 156     for(iidx=0;iidx<4*DIM;iidx++)
 157     {
 158         scratch[iidx] = 0.0;
 159     }
 160
 161     /* Start outer loop over neighborlists */
 162     for(iidx=0; iidx<nri; iidx++)
 163     {
 164         /* Load shift vector for this list */
 165         i_shift_offset   = DIM*shiftidx[iidx];
 166
 167         /* Load limits for loop over neighbors */
 168         j_index_start    = jindex[iidx];
 169         j_index_end      = jindex[iidx+1];
 170
 171         /* Get outer coordinate index */
 172         inr              = iinr[iidx];
 173         i_coord_offset   = DIM*inr;
 174
 175         /* Load i particle coords and add shift vector */
 176         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 177                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 178
 179         fix0             = _mm_setzero_ps();
 180         fiy0             = _mm_setzero_ps();
 181         fiz0             = _mm_setzero_ps();
 182         fix1             = _mm_setzero_ps();
 183         fiy1             = _mm_setzero_ps();
 184         fiz1             = _mm_setzero_ps();
 185         fix2             = _mm_setzero_ps();
 186         fiy2             = _mm_setzero_ps();
 187         fiz2             = _mm_setzero_ps();
 188
 189         /* Reset potential sums */
 190         velecsum         = _mm_setzero_ps();
 191
 192         /* Start inner kernel loop */
 193         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 194         {
 195
 196             /* Get j neighbor index, and coordinate index */
 197             jnrA             = jjnr[jidx];
 198             jnrB             = jjnr[jidx+1];
 199             jnrC             = jjnr[jidx+2];
 200             jnrD             = jjnr[jidx+3];
 201             j_coord_offsetA  = DIM*jnrA;
 202             j_coord_offsetB  = DIM*jnrB;
 203             j_coord_offsetC  = DIM*jnrC;
 204             j_coord_offsetD  = DIM*jnrD;
 205
 206             /* load j atom coordinates */
 207             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 208                                               x+j_coord_offsetC,x+j_coord_offsetD,
 209                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 210
 211             /* Calculate displacement vector */
 212             dx00             = _mm_sub_ps(ix0,jx0);
 213             dy00             = _mm_sub_ps(iy0,jy0);
 214             dz00             = _mm_sub_ps(iz0,jz0);
 215             dx01             = _mm_sub_ps(ix0,jx1);
 216             dy01             = _mm_sub_ps(iy0,jy1);
 217             dz01             = _mm_sub_ps(iz0,jz1);
 218             dx02             = _mm_sub_ps(ix0,jx2);
 219             dy02             = _mm_sub_ps(iy0,jy2);
 220             dz02             = _mm_sub_ps(iz0,jz2);
 221             dx10             = _mm_sub_ps(ix1,jx0);
 222             dy10             = _mm_sub_ps(iy1,jy0);
 223             dz10             = _mm_sub_ps(iz1,jz0);
 224             dx11             = _mm_sub_ps(ix1,jx1);
 225             dy11             = _mm_sub_ps(iy1,jy1);
 226             dz11             = _mm_sub_ps(iz1,jz1);
 227             dx12             = _mm_sub_ps(ix1,jx2);
 228             dy12             = _mm_sub_ps(iy1,jy2);
 229             dz12             = _mm_sub_ps(iz1,jz2);
 230             dx20             = _mm_sub_ps(ix2,jx0);
 231             dy20             = _mm_sub_ps(iy2,jy0);
 232             dz20             = _mm_sub_ps(iz2,jz0);
 233             dx21             = _mm_sub_ps(ix2,jx1);
 234             dy21             = _mm_sub_ps(iy2,jy1);
 235             dz21             = _mm_sub_ps(iz2,jz1);
 236             dx22             = _mm_sub_ps(ix2,jx2);
 237             dy22             = _mm_sub_ps(iy2,jy2);
 238             dz22             = _mm_sub_ps(iz2,jz2);
 239
 240             /* Calculate squared distance and things based on it */
 241             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 242             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 243             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 244             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 245             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 246             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 247             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 248             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 249             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 250
 251             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 252             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 253             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 254             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 255             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 256             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 257             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 258             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 259             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 260
 261             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 262             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 263             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 264             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 265             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 266             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 267             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 268             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 269             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 270
 271             fjx0             = _mm_setzero_ps();
 272             fjy0             = _mm_setzero_ps();
 273             fjz0             = _mm_setzero_ps();
 274             fjx1             = _mm_setzero_ps();
 275             fjy1             = _mm_setzero_ps();
 276             fjz1             = _mm_setzero_ps();
 277             fjx2             = _mm_setzero_ps();
 278             fjy2             = _mm_setzero_ps();
 279             fjz2             = _mm_setzero_ps();
 280
 281             /**************************
 282              * CALCULATE INTERACTIONS *
 283              **************************/
 284
 285             /* REACTION-FIELD ELECTROSTATICS */
 286             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_add_ps(rinv00,_mm_mul_ps(krf,rsq00)),crf));
 287             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
 288
 289             /* Update potential sum for this i atom from the interaction with this j atom. */
 290             velecsum         = _mm_add_ps(velecsum,velec);
 291
 292             fscal            = felec;
 293
 294             /* Calculate temporary vectorial force */
 295             tx               = _mm_mul_ps(fscal,dx00);
 296             ty               = _mm_mul_ps(fscal,dy00);
 297             tz               = _mm_mul_ps(fscal,dz00);
 298
 299             /* Update vectorial force */
 300             fix0             = _mm_add_ps(fix0,tx);
 301             fiy0             = _mm_add_ps(fiy0,ty);
 302             fiz0             = _mm_add_ps(fiz0,tz);
 303
 304             fjx0             = _mm_add_ps(fjx0,tx);
 305             fjy0             = _mm_add_ps(fjy0,ty);
 306             fjz0             = _mm_add_ps(fjz0,tz);
 307
 308             /**************************
 309              * CALCULATE INTERACTIONS *
 310              **************************/
 311
 312             /* REACTION-FIELD ELECTROSTATICS */
 313             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_add_ps(rinv01,_mm_mul_ps(krf,rsq01)),crf));
 314             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
 315
 316             /* Update potential sum for this i atom from the interaction with this j atom. */
 317             velecsum         = _mm_add_ps(velecsum,velec);
 318
 319             fscal            = felec;
 320
 321             /* Calculate temporary vectorial force */
 322             tx               = _mm_mul_ps(fscal,dx01);
 323             ty               = _mm_mul_ps(fscal,dy01);
 324             tz               = _mm_mul_ps(fscal,dz01);
 325
 326             /* Update vectorial force */
 327             fix0             = _mm_add_ps(fix0,tx);
 328             fiy0             = _mm_add_ps(fiy0,ty);
 329             fiz0             = _mm_add_ps(fiz0,tz);
 330
 331             fjx1             = _mm_add_ps(fjx1,tx);
 332             fjy1             = _mm_add_ps(fjy1,ty);
 333             fjz1             = _mm_add_ps(fjz1,tz);
 334
 335             /**************************
 336              * CALCULATE INTERACTIONS *
 337              **************************/
 338
 339             /* REACTION-FIELD ELECTROSTATICS */
 340             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_add_ps(rinv02,_mm_mul_ps(krf,rsq02)),crf));
 341             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
 342
 343             /* Update potential sum for this i atom from the interaction with this j atom. */
 344             velecsum         = _mm_add_ps(velecsum,velec);
 345
 346             fscal            = felec;
 347
 348             /* Calculate temporary vectorial force */
 349             tx               = _mm_mul_ps(fscal,dx02);
 350             ty               = _mm_mul_ps(fscal,dy02);
 351             tz               = _mm_mul_ps(fscal,dz02);
 352
 353             /* Update vectorial force */
 354             fix0             = _mm_add_ps(fix0,tx);
 355             fiy0             = _mm_add_ps(fiy0,ty);
 356             fiz0             = _mm_add_ps(fiz0,tz);
 357
 358             fjx2             = _mm_add_ps(fjx2,tx);
 359             fjy2             = _mm_add_ps(fjy2,ty);
 360             fjz2             = _mm_add_ps(fjz2,tz);
 361
 362             /**************************
 363              * CALCULATE INTERACTIONS *
 364              **************************/
 365
 366             /* REACTION-FIELD ELECTROSTATICS */
 367             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_add_ps(rinv10,_mm_mul_ps(krf,rsq10)),crf));
 368             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
 369
 370             /* Update potential sum for this i atom from the interaction with this j atom. */
 371             velecsum         = _mm_add_ps(velecsum,velec);
 372
 373             fscal            = felec;
 374
 375             /* Calculate temporary vectorial force */
 376             tx               = _mm_mul_ps(fscal,dx10);
 377             ty               = _mm_mul_ps(fscal,dy10);
 378             tz               = _mm_mul_ps(fscal,dz10);
 379
 380             /* Update vectorial force */
 381             fix1             = _mm_add_ps(fix1,tx);
 382             fiy1             = _mm_add_ps(fiy1,ty);
 383             fiz1             = _mm_add_ps(fiz1,tz);
 384
 385             fjx0             = _mm_add_ps(fjx0,tx);
 386             fjy0             = _mm_add_ps(fjy0,ty);
 387             fjz0             = _mm_add_ps(fjz0,tz);
 388
 389             /**************************
 390              * CALCULATE INTERACTIONS *
 391              **************************/
 392
 393             /* REACTION-FIELD ELECTROSTATICS */
 394             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 395             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 396
 397             /* Update potential sum for this i atom from the interaction with this j atom. */
 398             velecsum         = _mm_add_ps(velecsum,velec);
 399
 400             fscal            = felec;
 401
 402             /* Calculate temporary vectorial force */
 403             tx               = _mm_mul_ps(fscal,dx11);
 404             ty               = _mm_mul_ps(fscal,dy11);
 405             tz               = _mm_mul_ps(fscal,dz11);
 406
 407             /* Update vectorial force */
 408             fix1             = _mm_add_ps(fix1,tx);
 409             fiy1             = _mm_add_ps(fiy1,ty);
 410             fiz1             = _mm_add_ps(fiz1,tz);
 411
 412             fjx1             = _mm_add_ps(fjx1,tx);
 413             fjy1             = _mm_add_ps(fjy1,ty);
 414             fjz1             = _mm_add_ps(fjz1,tz);
 415
 416             /**************************
 417              * CALCULATE INTERACTIONS *
 418              **************************/
 419
 420             /* REACTION-FIELD ELECTROSTATICS */
 421             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 422             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 423
 424             /* Update potential sum for this i atom from the interaction with this j atom. */
 425             velecsum         = _mm_add_ps(velecsum,velec);
 426
 427             fscal            = felec;
 428
 429             /* Calculate temporary vectorial force */
 430             tx               = _mm_mul_ps(fscal,dx12);
 431             ty               = _mm_mul_ps(fscal,dy12);
 432             tz               = _mm_mul_ps(fscal,dz12);
 433
 434             /* Update vectorial force */
 435             fix1             = _mm_add_ps(fix1,tx);
 436             fiy1             = _mm_add_ps(fiy1,ty);
 437             fiz1             = _mm_add_ps(fiz1,tz);
 438
 439             fjx2             = _mm_add_ps(fjx2,tx);
 440             fjy2             = _mm_add_ps(fjy2,ty);
 441             fjz2             = _mm_add_ps(fjz2,tz);
 442
 443             /**************************
 444              * CALCULATE INTERACTIONS *
 445              **************************/
 446
 447             /* REACTION-FIELD ELECTROSTATICS */
 448             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_add_ps(rinv20,_mm_mul_ps(krf,rsq20)),crf));
 449             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
 450
 451             /* Update potential sum for this i atom from the interaction with this j atom. */
 452             velecsum         = _mm_add_ps(velecsum,velec);
 453
 454             fscal            = felec;
 455
 456             /* Calculate temporary vectorial force */
 457             tx               = _mm_mul_ps(fscal,dx20);
 458             ty               = _mm_mul_ps(fscal,dy20);
 459             tz               = _mm_mul_ps(fscal,dz20);
 460
 461             /* Update vectorial force */
 462             fix2             = _mm_add_ps(fix2,tx);
 463             fiy2             = _mm_add_ps(fiy2,ty);
 464             fiz2             = _mm_add_ps(fiz2,tz);
 465
 466             fjx0             = _mm_add_ps(fjx0,tx);
 467             fjy0             = _mm_add_ps(fjy0,ty);
 468             fjz0             = _mm_add_ps(fjz0,tz);
 469
 470             /**************************
 471              * CALCULATE INTERACTIONS *
 472              **************************/
 473
 474             /* REACTION-FIELD ELECTROSTATICS */
 475             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 476             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 477
 478             /* Update potential sum for this i atom from the interaction with this j atom. */
 479             velecsum         = _mm_add_ps(velecsum,velec);
 480
 481             fscal            = felec;
 482
 483             /* Calculate temporary vectorial force */
 484             tx               = _mm_mul_ps(fscal,dx21);
 485             ty               = _mm_mul_ps(fscal,dy21);
 486             tz               = _mm_mul_ps(fscal,dz21);
 487
 488             /* Update vectorial force */
 489             fix2             = _mm_add_ps(fix2,tx);
 490             fiy2             = _mm_add_ps(fiy2,ty);
 491             fiz2             = _mm_add_ps(fiz2,tz);
 492
 493             fjx1             = _mm_add_ps(fjx1,tx);
 494             fjy1             = _mm_add_ps(fjy1,ty);
 495             fjz1             = _mm_add_ps(fjz1,tz);
 496
 497             /**************************
 498              * CALCULATE INTERACTIONS *
 499              **************************/
 500
 501             /* REACTION-FIELD ELECTROSTATICS */
 502             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 503             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 504
 505             /* Update potential sum for this i atom from the interaction with this j atom. */
 506             velecsum         = _mm_add_ps(velecsum,velec);
 507
 508             fscal            = felec;
 509
 510             /* Calculate temporary vectorial force */
 511             tx               = _mm_mul_ps(fscal,dx22);
 512             ty               = _mm_mul_ps(fscal,dy22);
 513             tz               = _mm_mul_ps(fscal,dz22);
 514
 515             /* Update vectorial force */
 516             fix2             = _mm_add_ps(fix2,tx);
 517             fiy2             = _mm_add_ps(fiy2,ty);
 518             fiz2             = _mm_add_ps(fiz2,tz);
 519
 520             fjx2             = _mm_add_ps(fjx2,tx);
 521             fjy2             = _mm_add_ps(fjy2,ty);
 522             fjz2             = _mm_add_ps(fjz2,tz);
 523
 524             fjptrA             = f+j_coord_offsetA;
 525             fjptrB             = f+j_coord_offsetB;
 526             fjptrC             = f+j_coord_offsetC;
 527             fjptrD             = f+j_coord_offsetD;
 528
 529             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 530                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 531
 532             /* Inner loop uses 288 flops */
 533         }
 534
 535         if(jidx<j_index_end)
 536         {
 537
 538             /* Get j neighbor index, and coordinate index */
 539             jnrlistA         = jjnr[jidx];
 540             jnrlistB         = jjnr[jidx+1];
 541             jnrlistC         = jjnr[jidx+2];
 542             jnrlistD         = jjnr[jidx+3];
 543             /* Sign of each element will be negative for non-real atoms.
 544              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 545              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 546              */
 547             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 548             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 549             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 550             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 551             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 552             j_coord_offsetA  = DIM*jnrA;
 553             j_coord_offsetB  = DIM*jnrB;
 554             j_coord_offsetC  = DIM*jnrC;
 555             j_coord_offsetD  = DIM*jnrD;
 556
 557             /* load j atom coordinates */
 558             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 559                                               x+j_coord_offsetC,x+j_coord_offsetD,
 560                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 561
 562             /* Calculate displacement vector */
 563             dx00             = _mm_sub_ps(ix0,jx0);
 564             dy00             = _mm_sub_ps(iy0,jy0);
 565             dz00             = _mm_sub_ps(iz0,jz0);
 566             dx01             = _mm_sub_ps(ix0,jx1);
 567             dy01             = _mm_sub_ps(iy0,jy1);
 568             dz01             = _mm_sub_ps(iz0,jz1);
 569             dx02             = _mm_sub_ps(ix0,jx2);
 570             dy02             = _mm_sub_ps(iy0,jy2);
 571             dz02             = _mm_sub_ps(iz0,jz2);
 572             dx10             = _mm_sub_ps(ix1,jx0);
 573             dy10             = _mm_sub_ps(iy1,jy0);
 574             dz10             = _mm_sub_ps(iz1,jz0);
 575             dx11             = _mm_sub_ps(ix1,jx1);
 576             dy11             = _mm_sub_ps(iy1,jy1);
 577             dz11             = _mm_sub_ps(iz1,jz1);
 578             dx12             = _mm_sub_ps(ix1,jx2);
 579             dy12             = _mm_sub_ps(iy1,jy2);
 580             dz12             = _mm_sub_ps(iz1,jz2);
 581             dx20             = _mm_sub_ps(ix2,jx0);
 582             dy20             = _mm_sub_ps(iy2,jy0);
 583             dz20             = _mm_sub_ps(iz2,jz0);
 584             dx21             = _mm_sub_ps(ix2,jx1);
 585             dy21             = _mm_sub_ps(iy2,jy1);
 586             dz21             = _mm_sub_ps(iz2,jz1);
 587             dx22             = _mm_sub_ps(ix2,jx2);
 588             dy22             = _mm_sub_ps(iy2,jy2);
 589             dz22             = _mm_sub_ps(iz2,jz2);
 590
 591             /* Calculate squared distance and things based on it */
 592             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 593             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 594             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 595             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 596             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 597             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 598             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 599             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 600             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 601
 602             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 603             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 604             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 605             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 606             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 607             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 608             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 609             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 610             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 611
 612             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 613             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 614             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 615             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 616             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 617             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 618             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 619             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 620             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 621
 622             fjx0             = _mm_setzero_ps();
 623             fjy0             = _mm_setzero_ps();
 624             fjz0             = _mm_setzero_ps();
 625             fjx1             = _mm_setzero_ps();
 626             fjy1             = _mm_setzero_ps();
 627             fjz1             = _mm_setzero_ps();
 628             fjx2             = _mm_setzero_ps();
 629             fjy2             = _mm_setzero_ps();
 630             fjz2             = _mm_setzero_ps();
 631
 632             /**************************
 633              * CALCULATE INTERACTIONS *
 634              **************************/
 635
 636             /* REACTION-FIELD ELECTROSTATICS */
 637             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_add_ps(rinv00,_mm_mul_ps(krf,rsq00)),crf));
 638             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
 639
 640             /* Update potential sum for this i atom from the interaction with this j atom. */
 641             velec            = _mm_andnot_ps(dummy_mask,velec);
 642             velecsum         = _mm_add_ps(velecsum,velec);
 643
 644             fscal            = felec;
 645
 646             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 647
 648             /* Calculate temporary vectorial force */
 649             tx               = _mm_mul_ps(fscal,dx00);
 650             ty               = _mm_mul_ps(fscal,dy00);
 651             tz               = _mm_mul_ps(fscal,dz00);
 652
 653             /* Update vectorial force */
 654             fix0             = _mm_add_ps(fix0,tx);
 655             fiy0             = _mm_add_ps(fiy0,ty);
 656             fiz0             = _mm_add_ps(fiz0,tz);
 657
 658             fjx0             = _mm_add_ps(fjx0,tx);
 659             fjy0             = _mm_add_ps(fjy0,ty);
 660             fjz0             = _mm_add_ps(fjz0,tz);
 661
 662             /**************************
 663              * CALCULATE INTERACTIONS *
 664              **************************/
 665
 666             /* REACTION-FIELD ELECTROSTATICS */
 667             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_add_ps(rinv01,_mm_mul_ps(krf,rsq01)),crf));
 668             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
 669
 670             /* Update potential sum for this i atom from the interaction with this j atom. */
 671             velec            = _mm_andnot_ps(dummy_mask,velec);
 672             velecsum         = _mm_add_ps(velecsum,velec);
 673
 674             fscal            = felec;
 675
 676             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 677
 678             /* Calculate temporary vectorial force */
 679             tx               = _mm_mul_ps(fscal,dx01);
 680             ty               = _mm_mul_ps(fscal,dy01);
 681             tz               = _mm_mul_ps(fscal,dz01);
 682
 683             /* Update vectorial force */
 684             fix0             = _mm_add_ps(fix0,tx);
 685             fiy0             = _mm_add_ps(fiy0,ty);
 686             fiz0             = _mm_add_ps(fiz0,tz);
 687
 688             fjx1             = _mm_add_ps(fjx1,tx);
 689             fjy1             = _mm_add_ps(fjy1,ty);
 690             fjz1             = _mm_add_ps(fjz1,tz);
 691
 692             /**************************
 693              * CALCULATE INTERACTIONS *
 694              **************************/
 695
 696             /* REACTION-FIELD ELECTROSTATICS */
 697             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_add_ps(rinv02,_mm_mul_ps(krf,rsq02)),crf));
 698             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
 699
 700             /* Update potential sum for this i atom from the interaction with this j atom. */
 701             velec            = _mm_andnot_ps(dummy_mask,velec);
 702             velecsum         = _mm_add_ps(velecsum,velec);
 703
 704             fscal            = felec;
 705
 706             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 707
 708             /* Calculate temporary vectorial force */
 709             tx               = _mm_mul_ps(fscal,dx02);
 710             ty               = _mm_mul_ps(fscal,dy02);
 711             tz               = _mm_mul_ps(fscal,dz02);
 712
 713             /* Update vectorial force */
 714             fix0             = _mm_add_ps(fix0,tx);
 715             fiy0             = _mm_add_ps(fiy0,ty);
 716             fiz0             = _mm_add_ps(fiz0,tz);
 717
 718             fjx2             = _mm_add_ps(fjx2,tx);
 719             fjy2             = _mm_add_ps(fjy2,ty);
 720             fjz2             = _mm_add_ps(fjz2,tz);
 721
 722             /**************************
 723              * CALCULATE INTERACTIONS *
 724              **************************/
 725
 726             /* REACTION-FIELD ELECTROSTATICS */
 727             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_add_ps(rinv10,_mm_mul_ps(krf,rsq10)),crf));
 728             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
 729
 730             /* Update potential sum for this i atom from the interaction with this j atom. */
 731             velec            = _mm_andnot_ps(dummy_mask,velec);
 732             velecsum         = _mm_add_ps(velecsum,velec);
 733
 734             fscal            = felec;
 735
 736             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 737
 738             /* Calculate temporary vectorial force */
 739             tx               = _mm_mul_ps(fscal,dx10);
 740             ty               = _mm_mul_ps(fscal,dy10);
 741             tz               = _mm_mul_ps(fscal,dz10);
 742
 743             /* Update vectorial force */
 744             fix1             = _mm_add_ps(fix1,tx);
 745             fiy1             = _mm_add_ps(fiy1,ty);
 746             fiz1             = _mm_add_ps(fiz1,tz);
 747
 748             fjx0             = _mm_add_ps(fjx0,tx);
 749             fjy0             = _mm_add_ps(fjy0,ty);
 750             fjz0             = _mm_add_ps(fjz0,tz);
 751
 752             /**************************
 753              * CALCULATE INTERACTIONS *
 754              **************************/
 755
 756             /* REACTION-FIELD ELECTROSTATICS */
 757             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_add_ps(rinv11,_mm_mul_ps(krf,rsq11)),crf));
 758             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
 759
 760             /* Update potential sum for this i atom from the interaction with this j atom. */
 761             velec            = _mm_andnot_ps(dummy_mask,velec);
 762             velecsum         = _mm_add_ps(velecsum,velec);
 763
 764             fscal            = felec;
 765
 766             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 767
 768             /* Calculate temporary vectorial force */
 769             tx               = _mm_mul_ps(fscal,dx11);
 770             ty               = _mm_mul_ps(fscal,dy11);
 771             tz               = _mm_mul_ps(fscal,dz11);
 772
 773             /* Update vectorial force */
 774             fix1             = _mm_add_ps(fix1,tx);
 775             fiy1             = _mm_add_ps(fiy1,ty);
 776             fiz1             = _mm_add_ps(fiz1,tz);
 777
 778             fjx1             = _mm_add_ps(fjx1,tx);
 779             fjy1             = _mm_add_ps(fjy1,ty);
 780             fjz1             = _mm_add_ps(fjz1,tz);
 781
 782             /**************************
 783              * CALCULATE INTERACTIONS *
 784              **************************/
 785
 786             /* REACTION-FIELD ELECTROSTATICS */
 787             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_add_ps(rinv12,_mm_mul_ps(krf,rsq12)),crf));
 788             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
 789
 790             /* Update potential sum for this i atom from the interaction with this j atom. */
 791             velec            = _mm_andnot_ps(dummy_mask,velec);
 792             velecsum         = _mm_add_ps(velecsum,velec);
 793
 794             fscal            = felec;
 795
 796             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 797
 798             /* Calculate temporary vectorial force */
 799             tx               = _mm_mul_ps(fscal,dx12);
 800             ty               = _mm_mul_ps(fscal,dy12);
 801             tz               = _mm_mul_ps(fscal,dz12);
 802
 803             /* Update vectorial force */
 804             fix1             = _mm_add_ps(fix1,tx);
 805             fiy1             = _mm_add_ps(fiy1,ty);
 806             fiz1             = _mm_add_ps(fiz1,tz);
 807
 808             fjx2             = _mm_add_ps(fjx2,tx);
 809             fjy2             = _mm_add_ps(fjy2,ty);
 810             fjz2             = _mm_add_ps(fjz2,tz);
 811
 812             /**************************
 813              * CALCULATE INTERACTIONS *
 814              **************************/
 815
 816             /* REACTION-FIELD ELECTROSTATICS */
 817             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_add_ps(rinv20,_mm_mul_ps(krf,rsq20)),crf));
 818             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
 819
 820             /* Update potential sum for this i atom from the interaction with this j atom. */
 821             velec            = _mm_andnot_ps(dummy_mask,velec);
 822             velecsum         = _mm_add_ps(velecsum,velec);
 823
 824             fscal            = felec;
 825
 826             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 827
 828             /* Calculate temporary vectorial force */
 829             tx               = _mm_mul_ps(fscal,dx20);
 830             ty               = _mm_mul_ps(fscal,dy20);
 831             tz               = _mm_mul_ps(fscal,dz20);
 832
 833             /* Update vectorial force */
 834             fix2             = _mm_add_ps(fix2,tx);
 835             fiy2             = _mm_add_ps(fiy2,ty);
 836             fiz2             = _mm_add_ps(fiz2,tz);
 837
 838             fjx0             = _mm_add_ps(fjx0,tx);
 839             fjy0             = _mm_add_ps(fjy0,ty);
 840             fjz0             = _mm_add_ps(fjz0,tz);
 841
 842             /**************************
 843              * CALCULATE INTERACTIONS *
 844              **************************/
 845
 846             /* REACTION-FIELD ELECTROSTATICS */
 847             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_add_ps(rinv21,_mm_mul_ps(krf,rsq21)),crf));
 848             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
 849
 850             /* Update potential sum for this i atom from the interaction with this j atom. */
 851             velec            = _mm_andnot_ps(dummy_mask,velec);
 852             velecsum         = _mm_add_ps(velecsum,velec);
 853
 854             fscal            = felec;
 855
 856             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 857
 858             /* Calculate temporary vectorial force */
 859             tx               = _mm_mul_ps(fscal,dx21);
 860             ty               = _mm_mul_ps(fscal,dy21);
 861             tz               = _mm_mul_ps(fscal,dz21);
 862
 863             /* Update vectorial force */
 864             fix2             = _mm_add_ps(fix2,tx);
 865             fiy2             = _mm_add_ps(fiy2,ty);
 866             fiz2             = _mm_add_ps(fiz2,tz);
 867
 868             fjx1             = _mm_add_ps(fjx1,tx);
 869             fjy1             = _mm_add_ps(fjy1,ty);
 870             fjz1             = _mm_add_ps(fjz1,tz);
 871
 872             /**************************
 873              * CALCULATE INTERACTIONS *
 874              **************************/
 875
 876             /* REACTION-FIELD ELECTROSTATICS */
 877             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_add_ps(rinv22,_mm_mul_ps(krf,rsq22)),crf));
 878             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
 879
 880             /* Update potential sum for this i atom from the interaction with this j atom. */
 881             velec            = _mm_andnot_ps(dummy_mask,velec);
 882             velecsum         = _mm_add_ps(velecsum,velec);
 883
 884             fscal            = felec;
 885
 886             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 887
 888             /* Calculate temporary vectorial force */
 889             tx               = _mm_mul_ps(fscal,dx22);
 890             ty               = _mm_mul_ps(fscal,dy22);
 891             tz               = _mm_mul_ps(fscal,dz22);
 892
 893             /* Update vectorial force */
 894             fix2             = _mm_add_ps(fix2,tx);
 895             fiy2             = _mm_add_ps(fiy2,ty);
 896             fiz2             = _mm_add_ps(fiz2,tz);
 897
 898             fjx2             = _mm_add_ps(fjx2,tx);
 899             fjy2             = _mm_add_ps(fjy2,ty);
 900             fjz2             = _mm_add_ps(fjz2,tz);
 901
 902             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 903             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 904             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 905             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 906
 907             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 908                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 909
 910             /* Inner loop uses 288 flops */
 911         }
 912
 913         /* End of innermost loop */
 914
 915         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 916                                               f+i_coord_offset,fshift+i_shift_offset);
 917
 918         ggid                        = gid[iidx];
 919         /* Update potential energies */
 920         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 921
 922         /* Increment number of inner iterations */
 923         inneriter                  += j_index_end - j_index_start;
 924
 925         /* Outer loop uses 19 flops */
 926     }
 927
 928     /* Increment number of outer iterations */
 929     outeriter        += nri;
 930
 931     /* Update outer/inner flops */
 932
 933     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_VF,outeriter*19 + inneriter*288);
 934 }
 935 /*
 936  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwNone_GeomW3W3_F_sse4_1_single
 937  * Electrostatics interaction: ReactionField
 938  * VdW interaction:            None
 939  * Geometry:                   Water3-Water3
 940  * Calculate force/pot:        Force
 941  */
 942 void
 943 nb_kernel_ElecRF_VdwNone_GeomW3W3_F_sse4_1_single
 944                     (t_nblist                    * gmx_restrict       nlist,
 945                      rvec                        * gmx_restrict          xx,
 946                      rvec                        * gmx_restrict          ff,
 947                      t_forcerec                  * gmx_restrict          fr,
 948                      t_mdatoms                   * gmx_restrict     mdatoms,
 949                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 950                      t_nrnb                      * gmx_restrict        nrnb)
 951 {
 952     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 953      * just 0 for non-waters.
 954      * Suffixes A,B,C,D refer to j loop unrolling done with SSE, e.g. for the four different
 955      * jnr indices corresponding to data put in the four positions in the SIMD register.
 956      */
 957     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 958     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 959     int              jnrA,jnrB,jnrC,jnrD;
 960     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 961     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 962     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 963     real             rcutoff_scalar;
 964     real             *shiftvec,*fshift,*x,*f;
 965     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 966     real             scratch[4*DIM];
 967     __m128           tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 968     int              vdwioffset0;
 969     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 970     int              vdwioffset1;
 971     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 972     int              vdwioffset2;
 973     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 974     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 975     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 976     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
 977     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 978     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
 979     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 980     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 981     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
 982     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 983     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 984     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 985     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 986     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 987     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 988     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 989     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 990     real             *charge;
 991     __m128           dummy_mask,cutoff_mask;
 992     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 993     __m128           one     = _mm_set1_ps(1.0);
 994     __m128           two     = _mm_set1_ps(2.0);
 995     x                = xx[0];
 996     f                = ff[0];
 997
 998     nri              = nlist->nri;
 999     iinr             = nlist->iinr;
1000     jindex           = nlist->jindex;
1001     jjnr             = nlist->jjnr;
1002     shiftidx         = nlist->shift;
1003     gid              = nlist->gid;
1004     shiftvec         = fr->shift_vec[0];
1005     fshift           = fr->fshift[0];
1006     facel            = _mm_set1_ps(fr->epsfac);
1007     charge           = mdatoms->chargeA;
1008     krf              = _mm_set1_ps(fr->ic->k_rf);
1009     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1010     crf              = _mm_set1_ps(fr->ic->c_rf);
1011
1012     /* Setup water-specific parameters */
1013     inr              = nlist->iinr[0];
1014     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1015     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1016     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1017
1018     jq0              = _mm_set1_ps(charge[inr+0]);
1019     jq1              = _mm_set1_ps(charge[inr+1]);
1020     jq2              = _mm_set1_ps(charge[inr+2]);
1021     qq00             = _mm_mul_ps(iq0,jq0);
1022     qq01             = _mm_mul_ps(iq0,jq1);
1023     qq02             = _mm_mul_ps(iq0,jq2);
1024     qq10             = _mm_mul_ps(iq1,jq0);
1025     qq11             = _mm_mul_ps(iq1,jq1);
1026     qq12             = _mm_mul_ps(iq1,jq2);
1027     qq20             = _mm_mul_ps(iq2,jq0);
1028     qq21             = _mm_mul_ps(iq2,jq1);
1029     qq22             = _mm_mul_ps(iq2,jq2);
1030
1031     /* Avoid stupid compiler warnings */
1032     jnrA = jnrB = jnrC = jnrD = 0;
1033     j_coord_offsetA = 0;
1034     j_coord_offsetB = 0;
1035     j_coord_offsetC = 0;
1036     j_coord_offsetD = 0;
1037
1038     outeriter        = 0;
1039     inneriter        = 0;
1040
1041     for(iidx=0;iidx<4*DIM;iidx++)
1042     {
1043         scratch[iidx] = 0.0;
1044     }
1045
1046     /* Start outer loop over neighborlists */
1047     for(iidx=0; iidx<nri; iidx++)
1048     {
1049         /* Load shift vector for this list */
1050         i_shift_offset   = DIM*shiftidx[iidx];
1051
1052         /* Load limits for loop over neighbors */
1053         j_index_start    = jindex[iidx];
1054         j_index_end      = jindex[iidx+1];
1055
1056         /* Get outer coordinate index */
1057         inr              = iinr[iidx];
1058         i_coord_offset   = DIM*inr;
1059
1060         /* Load i particle coords and add shift vector */
1061         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1062                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1063
1064         fix0             = _mm_setzero_ps();
1065         fiy0             = _mm_setzero_ps();
1066         fiz0             = _mm_setzero_ps();
1067         fix1             = _mm_setzero_ps();
1068         fiy1             = _mm_setzero_ps();
1069         fiz1             = _mm_setzero_ps();
1070         fix2             = _mm_setzero_ps();
1071         fiy2             = _mm_setzero_ps();
1072         fiz2             = _mm_setzero_ps();
1073
1074         /* Start inner kernel loop */
1075         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1076         {
1077
1078             /* Get j neighbor index, and coordinate index */
1079             jnrA             = jjnr[jidx];
1080             jnrB             = jjnr[jidx+1];
1081             jnrC             = jjnr[jidx+2];
1082             jnrD             = jjnr[jidx+3];
1083             j_coord_offsetA  = DIM*jnrA;
1084             j_coord_offsetB  = DIM*jnrB;
1085             j_coord_offsetC  = DIM*jnrC;
1086             j_coord_offsetD  = DIM*jnrD;
1087
1088             /* load j atom coordinates */
1089             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1090                                               x+j_coord_offsetC,x+j_coord_offsetD,
1091                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1092
1093             /* Calculate displacement vector */
1094             dx00             = _mm_sub_ps(ix0,jx0);
1095             dy00             = _mm_sub_ps(iy0,jy0);
1096             dz00             = _mm_sub_ps(iz0,jz0);
1097             dx01             = _mm_sub_ps(ix0,jx1);
1098             dy01             = _mm_sub_ps(iy0,jy1);
1099             dz01             = _mm_sub_ps(iz0,jz1);
1100             dx02             = _mm_sub_ps(ix0,jx2);
1101             dy02             = _mm_sub_ps(iy0,jy2);
1102             dz02             = _mm_sub_ps(iz0,jz2);
1103             dx10             = _mm_sub_ps(ix1,jx0);
1104             dy10             = _mm_sub_ps(iy1,jy0);
1105             dz10             = _mm_sub_ps(iz1,jz0);
1106             dx11             = _mm_sub_ps(ix1,jx1);
1107             dy11             = _mm_sub_ps(iy1,jy1);
1108             dz11             = _mm_sub_ps(iz1,jz1);
1109             dx12             = _mm_sub_ps(ix1,jx2);
1110             dy12             = _mm_sub_ps(iy1,jy2);
1111             dz12             = _mm_sub_ps(iz1,jz2);
1112             dx20             = _mm_sub_ps(ix2,jx0);
1113             dy20             = _mm_sub_ps(iy2,jy0);
1114             dz20             = _mm_sub_ps(iz2,jz0);
1115             dx21             = _mm_sub_ps(ix2,jx1);
1116             dy21             = _mm_sub_ps(iy2,jy1);
1117             dz21             = _mm_sub_ps(iz2,jz1);
1118             dx22             = _mm_sub_ps(ix2,jx2);
1119             dy22             = _mm_sub_ps(iy2,jy2);
1120             dz22             = _mm_sub_ps(iz2,jz2);
1121
1122             /* Calculate squared distance and things based on it */
1123             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1124             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1125             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1126             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1127             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1128             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1129             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1130             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1131             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1132
1133             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1134             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1135             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1136             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1137             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1138             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1139             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1140             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1141             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1142
1143             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1144             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1145             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1146             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1147             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1148             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1149             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1150             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1151             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1152
1153             fjx0             = _mm_setzero_ps();
1154             fjy0             = _mm_setzero_ps();
1155             fjz0             = _mm_setzero_ps();
1156             fjx1             = _mm_setzero_ps();
1157             fjy1             = _mm_setzero_ps();
1158             fjz1             = _mm_setzero_ps();
1159             fjx2             = _mm_setzero_ps();
1160             fjy2             = _mm_setzero_ps();
1161             fjz2             = _mm_setzero_ps();
1162
1163             /**************************
1164              * CALCULATE INTERACTIONS *
1165              **************************/
1166
1167             /* REACTION-FIELD ELECTROSTATICS */
1168             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
1169
1170             fscal            = felec;
1171
1172             /* Calculate temporary vectorial force */
1173             tx               = _mm_mul_ps(fscal,dx00);
1174             ty               = _mm_mul_ps(fscal,dy00);
1175             tz               = _mm_mul_ps(fscal,dz00);
1176
1177             /* Update vectorial force */
1178             fix0             = _mm_add_ps(fix0,tx);
1179             fiy0             = _mm_add_ps(fiy0,ty);
1180             fiz0             = _mm_add_ps(fiz0,tz);
1181
1182             fjx0             = _mm_add_ps(fjx0,tx);
1183             fjy0             = _mm_add_ps(fjy0,ty);
1184             fjz0             = _mm_add_ps(fjz0,tz);
1185
1186             /**************************
1187              * CALCULATE INTERACTIONS *
1188              **************************/
1189
1190             /* REACTION-FIELD ELECTROSTATICS */
1191             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
1192
1193             fscal            = felec;
1194
1195             /* Calculate temporary vectorial force */
1196             tx               = _mm_mul_ps(fscal,dx01);
1197             ty               = _mm_mul_ps(fscal,dy01);
1198             tz               = _mm_mul_ps(fscal,dz01);
1199
1200             /* Update vectorial force */
1201             fix0             = _mm_add_ps(fix0,tx);
1202             fiy0             = _mm_add_ps(fiy0,ty);
1203             fiz0             = _mm_add_ps(fiz0,tz);
1204
1205             fjx1             = _mm_add_ps(fjx1,tx);
1206             fjy1             = _mm_add_ps(fjy1,ty);
1207             fjz1             = _mm_add_ps(fjz1,tz);
1208
1209             /**************************
1210              * CALCULATE INTERACTIONS *
1211              **************************/
1212
1213             /* REACTION-FIELD ELECTROSTATICS */
1214             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
1215
1216             fscal            = felec;
1217
1218             /* Calculate temporary vectorial force */
1219             tx               = _mm_mul_ps(fscal,dx02);
1220             ty               = _mm_mul_ps(fscal,dy02);
1221             tz               = _mm_mul_ps(fscal,dz02);
1222
1223             /* Update vectorial force */
1224             fix0             = _mm_add_ps(fix0,tx);
1225             fiy0             = _mm_add_ps(fiy0,ty);
1226             fiz0             = _mm_add_ps(fiz0,tz);
1227
1228             fjx2             = _mm_add_ps(fjx2,tx);
1229             fjy2             = _mm_add_ps(fjy2,ty);
1230             fjz2             = _mm_add_ps(fjz2,tz);
1231
1232             /**************************
1233              * CALCULATE INTERACTIONS *
1234              **************************/
1235
1236             /* REACTION-FIELD ELECTROSTATICS */
1237             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
1238
1239             fscal            = felec;
1240
1241             /* Calculate temporary vectorial force */
1242             tx               = _mm_mul_ps(fscal,dx10);
1243             ty               = _mm_mul_ps(fscal,dy10);
1244             tz               = _mm_mul_ps(fscal,dz10);
1245
1246             /* Update vectorial force */
1247             fix1             = _mm_add_ps(fix1,tx);
1248             fiy1             = _mm_add_ps(fiy1,ty);
1249             fiz1             = _mm_add_ps(fiz1,tz);
1250
1251             fjx0             = _mm_add_ps(fjx0,tx);
1252             fjy0             = _mm_add_ps(fjy0,ty);
1253             fjz0             = _mm_add_ps(fjz0,tz);
1254
1255             /**************************
1256              * CALCULATE INTERACTIONS *
1257              **************************/
1258
1259             /* REACTION-FIELD ELECTROSTATICS */
1260             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1261
1262             fscal            = felec;
1263
1264             /* Calculate temporary vectorial force */
1265             tx               = _mm_mul_ps(fscal,dx11);
1266             ty               = _mm_mul_ps(fscal,dy11);
1267             tz               = _mm_mul_ps(fscal,dz11);
1268
1269             /* Update vectorial force */
1270             fix1             = _mm_add_ps(fix1,tx);
1271             fiy1             = _mm_add_ps(fiy1,ty);
1272             fiz1             = _mm_add_ps(fiz1,tz);
1273
1274             fjx1             = _mm_add_ps(fjx1,tx);
1275             fjy1             = _mm_add_ps(fjy1,ty);
1276             fjz1             = _mm_add_ps(fjz1,tz);
1277
1278             /**************************
1279              * CALCULATE INTERACTIONS *
1280              **************************/
1281
1282             /* REACTION-FIELD ELECTROSTATICS */
1283             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1284
1285             fscal            = felec;
1286
1287             /* Calculate temporary vectorial force */
1288             tx               = _mm_mul_ps(fscal,dx12);
1289             ty               = _mm_mul_ps(fscal,dy12);
1290             tz               = _mm_mul_ps(fscal,dz12);
1291
1292             /* Update vectorial force */
1293             fix1             = _mm_add_ps(fix1,tx);
1294             fiy1             = _mm_add_ps(fiy1,ty);
1295             fiz1             = _mm_add_ps(fiz1,tz);
1296
1297             fjx2             = _mm_add_ps(fjx2,tx);
1298             fjy2             = _mm_add_ps(fjy2,ty);
1299             fjz2             = _mm_add_ps(fjz2,tz);
1300
1301             /**************************
1302              * CALCULATE INTERACTIONS *
1303              **************************/
1304
1305             /* REACTION-FIELD ELECTROSTATICS */
1306             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
1307
1308             fscal            = felec;
1309
1310             /* Calculate temporary vectorial force */
1311             tx               = _mm_mul_ps(fscal,dx20);
1312             ty               = _mm_mul_ps(fscal,dy20);
1313             tz               = _mm_mul_ps(fscal,dz20);
1314
1315             /* Update vectorial force */
1316             fix2             = _mm_add_ps(fix2,tx);
1317             fiy2             = _mm_add_ps(fiy2,ty);
1318             fiz2             = _mm_add_ps(fiz2,tz);
1319
1320             fjx0             = _mm_add_ps(fjx0,tx);
1321             fjy0             = _mm_add_ps(fjy0,ty);
1322             fjz0             = _mm_add_ps(fjz0,tz);
1323
1324             /**************************
1325              * CALCULATE INTERACTIONS *
1326              **************************/
1327
1328             /* REACTION-FIELD ELECTROSTATICS */
1329             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1330
1331             fscal            = felec;
1332
1333             /* Calculate temporary vectorial force */
1334             tx               = _mm_mul_ps(fscal,dx21);
1335             ty               = _mm_mul_ps(fscal,dy21);
1336             tz               = _mm_mul_ps(fscal,dz21);
1337
1338             /* Update vectorial force */
1339             fix2             = _mm_add_ps(fix2,tx);
1340             fiy2             = _mm_add_ps(fiy2,ty);
1341             fiz2             = _mm_add_ps(fiz2,tz);
1342
1343             fjx1             = _mm_add_ps(fjx1,tx);
1344             fjy1             = _mm_add_ps(fjy1,ty);
1345             fjz1             = _mm_add_ps(fjz1,tz);
1346
1347             /**************************
1348              * CALCULATE INTERACTIONS *
1349              **************************/
1350
1351             /* REACTION-FIELD ELECTROSTATICS */
1352             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1353
1354             fscal            = felec;
1355
1356             /* Calculate temporary vectorial force */
1357             tx               = _mm_mul_ps(fscal,dx22);
1358             ty               = _mm_mul_ps(fscal,dy22);
1359             tz               = _mm_mul_ps(fscal,dz22);
1360
1361             /* Update vectorial force */
1362             fix2             = _mm_add_ps(fix2,tx);
1363             fiy2             = _mm_add_ps(fiy2,ty);
1364             fiz2             = _mm_add_ps(fiz2,tz);
1365
1366             fjx2             = _mm_add_ps(fjx2,tx);
1367             fjy2             = _mm_add_ps(fjy2,ty);
1368             fjz2             = _mm_add_ps(fjz2,tz);
1369
1370             fjptrA             = f+j_coord_offsetA;
1371             fjptrB             = f+j_coord_offsetB;
1372             fjptrC             = f+j_coord_offsetC;
1373             fjptrD             = f+j_coord_offsetD;
1374
1375             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1376                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1377
1378             /* Inner loop uses 243 flops */
1379         }
1380
1381         if(jidx<j_index_end)
1382         {
1383
1384             /* Get j neighbor index, and coordinate index */
1385             jnrlistA         = jjnr[jidx];
1386             jnrlistB         = jjnr[jidx+1];
1387             jnrlistC         = jjnr[jidx+2];
1388             jnrlistD         = jjnr[jidx+3];
1389             /* Sign of each element will be negative for non-real atoms.
1390              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1391              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1392              */
1393             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1394             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1395             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1396             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1397             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1398             j_coord_offsetA  = DIM*jnrA;
1399             j_coord_offsetB  = DIM*jnrB;
1400             j_coord_offsetC  = DIM*jnrC;
1401             j_coord_offsetD  = DIM*jnrD;
1402
1403             /* load j atom coordinates */
1404             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1405                                               x+j_coord_offsetC,x+j_coord_offsetD,
1406                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1407
1408             /* Calculate displacement vector */
1409             dx00             = _mm_sub_ps(ix0,jx0);
1410             dy00             = _mm_sub_ps(iy0,jy0);
1411             dz00             = _mm_sub_ps(iz0,jz0);
1412             dx01             = _mm_sub_ps(ix0,jx1);
1413             dy01             = _mm_sub_ps(iy0,jy1);
1414             dz01             = _mm_sub_ps(iz0,jz1);
1415             dx02             = _mm_sub_ps(ix0,jx2);
1416             dy02             = _mm_sub_ps(iy0,jy2);
1417             dz02             = _mm_sub_ps(iz0,jz2);
1418             dx10             = _mm_sub_ps(ix1,jx0);
1419             dy10             = _mm_sub_ps(iy1,jy0);
1420             dz10             = _mm_sub_ps(iz1,jz0);
1421             dx11             = _mm_sub_ps(ix1,jx1);
1422             dy11             = _mm_sub_ps(iy1,jy1);
1423             dz11             = _mm_sub_ps(iz1,jz1);
1424             dx12             = _mm_sub_ps(ix1,jx2);
1425             dy12             = _mm_sub_ps(iy1,jy2);
1426             dz12             = _mm_sub_ps(iz1,jz2);
1427             dx20             = _mm_sub_ps(ix2,jx0);
1428             dy20             = _mm_sub_ps(iy2,jy0);
1429             dz20             = _mm_sub_ps(iz2,jz0);
1430             dx21             = _mm_sub_ps(ix2,jx1);
1431             dy21             = _mm_sub_ps(iy2,jy1);
1432             dz21             = _mm_sub_ps(iz2,jz1);
1433             dx22             = _mm_sub_ps(ix2,jx2);
1434             dy22             = _mm_sub_ps(iy2,jy2);
1435             dz22             = _mm_sub_ps(iz2,jz2);
1436
1437             /* Calculate squared distance and things based on it */
1438             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1439             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1440             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1441             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1442             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1443             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1444             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1445             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1446             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1447
1448             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1449             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1450             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1451             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1452             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1453             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1454             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1455             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1456             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1457
1458             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1459             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1460             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1461             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1462             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1463             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1464             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1465             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1466             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1467
1468             fjx0             = _mm_setzero_ps();
1469             fjy0             = _mm_setzero_ps();
1470             fjz0             = _mm_setzero_ps();
1471             fjx1             = _mm_setzero_ps();
1472             fjy1             = _mm_setzero_ps();
1473             fjz1             = _mm_setzero_ps();
1474             fjx2             = _mm_setzero_ps();
1475             fjy2             = _mm_setzero_ps();
1476             fjz2             = _mm_setzero_ps();
1477
1478             /**************************
1479              * CALCULATE INTERACTIONS *
1480              **************************/
1481
1482             /* REACTION-FIELD ELECTROSTATICS */
1483             felec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_mul_ps(rinv00,rinvsq00),krf2));
1484
1485             fscal            = felec;
1486
1487             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1488
1489             /* Calculate temporary vectorial force */
1490             tx               = _mm_mul_ps(fscal,dx00);
1491             ty               = _mm_mul_ps(fscal,dy00);
1492             tz               = _mm_mul_ps(fscal,dz00);
1493
1494             /* Update vectorial force */
1495             fix0             = _mm_add_ps(fix0,tx);
1496             fiy0             = _mm_add_ps(fiy0,ty);
1497             fiz0             = _mm_add_ps(fiz0,tz);
1498
1499             fjx0             = _mm_add_ps(fjx0,tx);
1500             fjy0             = _mm_add_ps(fjy0,ty);
1501             fjz0             = _mm_add_ps(fjz0,tz);
1502
1503             /**************************
1504              * CALCULATE INTERACTIONS *
1505              **************************/
1506
1507             /* REACTION-FIELD ELECTROSTATICS */
1508             felec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_mul_ps(rinv01,rinvsq01),krf2));
1509
1510             fscal            = felec;
1511
1512             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1513
1514             /* Calculate temporary vectorial force */
1515             tx               = _mm_mul_ps(fscal,dx01);
1516             ty               = _mm_mul_ps(fscal,dy01);
1517             tz               = _mm_mul_ps(fscal,dz01);
1518
1519             /* Update vectorial force */
1520             fix0             = _mm_add_ps(fix0,tx);
1521             fiy0             = _mm_add_ps(fiy0,ty);
1522             fiz0             = _mm_add_ps(fiz0,tz);
1523
1524             fjx1             = _mm_add_ps(fjx1,tx);
1525             fjy1             = _mm_add_ps(fjy1,ty);
1526             fjz1             = _mm_add_ps(fjz1,tz);
1527
1528             /**************************
1529              * CALCULATE INTERACTIONS *
1530              **************************/
1531
1532             /* REACTION-FIELD ELECTROSTATICS */
1533             felec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_mul_ps(rinv02,rinvsq02),krf2));
1534
1535             fscal            = felec;
1536
1537             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1538
1539             /* Calculate temporary vectorial force */
1540             tx               = _mm_mul_ps(fscal,dx02);
1541             ty               = _mm_mul_ps(fscal,dy02);
1542             tz               = _mm_mul_ps(fscal,dz02);
1543
1544             /* Update vectorial force */
1545             fix0             = _mm_add_ps(fix0,tx);
1546             fiy0             = _mm_add_ps(fiy0,ty);
1547             fiz0             = _mm_add_ps(fiz0,tz);
1548
1549             fjx2             = _mm_add_ps(fjx2,tx);
1550             fjy2             = _mm_add_ps(fjy2,ty);
1551             fjz2             = _mm_add_ps(fjz2,tz);
1552
1553             /**************************
1554              * CALCULATE INTERACTIONS *
1555              **************************/
1556
1557             /* REACTION-FIELD ELECTROSTATICS */
1558             felec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_mul_ps(rinv10,rinvsq10),krf2));
1559
1560             fscal            = felec;
1561
1562             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1563
1564             /* Calculate temporary vectorial force */
1565             tx               = _mm_mul_ps(fscal,dx10);
1566             ty               = _mm_mul_ps(fscal,dy10);
1567             tz               = _mm_mul_ps(fscal,dz10);
1568
1569             /* Update vectorial force */
1570             fix1             = _mm_add_ps(fix1,tx);
1571             fiy1             = _mm_add_ps(fiy1,ty);
1572             fiz1             = _mm_add_ps(fiz1,tz);
1573
1574             fjx0             = _mm_add_ps(fjx0,tx);
1575             fjy0             = _mm_add_ps(fjy0,ty);
1576             fjz0             = _mm_add_ps(fjz0,tz);
1577
1578             /**************************
1579              * CALCULATE INTERACTIONS *
1580              **************************/
1581
1582             /* REACTION-FIELD ELECTROSTATICS */
1583             felec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_mul_ps(rinv11,rinvsq11),krf2));
1584
1585             fscal            = felec;
1586
1587             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1588
1589             /* Calculate temporary vectorial force */
1590             tx               = _mm_mul_ps(fscal,dx11);
1591             ty               = _mm_mul_ps(fscal,dy11);
1592             tz               = _mm_mul_ps(fscal,dz11);
1593
1594             /* Update vectorial force */
1595             fix1             = _mm_add_ps(fix1,tx);
1596             fiy1             = _mm_add_ps(fiy1,ty);
1597             fiz1             = _mm_add_ps(fiz1,tz);
1598
1599             fjx1             = _mm_add_ps(fjx1,tx);
1600             fjy1             = _mm_add_ps(fjy1,ty);
1601             fjz1             = _mm_add_ps(fjz1,tz);
1602
1603             /**************************
1604              * CALCULATE INTERACTIONS *
1605              **************************/
1606
1607             /* REACTION-FIELD ELECTROSTATICS */
1608             felec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_mul_ps(rinv12,rinvsq12),krf2));
1609
1610             fscal            = felec;
1611
1612             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1613
1614             /* Calculate temporary vectorial force */
1615             tx               = _mm_mul_ps(fscal,dx12);
1616             ty               = _mm_mul_ps(fscal,dy12);
1617             tz               = _mm_mul_ps(fscal,dz12);
1618
1619             /* Update vectorial force */
1620             fix1             = _mm_add_ps(fix1,tx);
1621             fiy1             = _mm_add_ps(fiy1,ty);
1622             fiz1             = _mm_add_ps(fiz1,tz);
1623
1624             fjx2             = _mm_add_ps(fjx2,tx);
1625             fjy2             = _mm_add_ps(fjy2,ty);
1626             fjz2             = _mm_add_ps(fjz2,tz);
1627
1628             /**************************
1629              * CALCULATE INTERACTIONS *
1630              **************************/
1631
1632             /* REACTION-FIELD ELECTROSTATICS */
1633             felec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_mul_ps(rinv20,rinvsq20),krf2));
1634
1635             fscal            = felec;
1636
1637             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1638
1639             /* Calculate temporary vectorial force */
1640             tx               = _mm_mul_ps(fscal,dx20);
1641             ty               = _mm_mul_ps(fscal,dy20);
1642             tz               = _mm_mul_ps(fscal,dz20);
1643
1644             /* Update vectorial force */
1645             fix2             = _mm_add_ps(fix2,tx);
1646             fiy2             = _mm_add_ps(fiy2,ty);
1647             fiz2             = _mm_add_ps(fiz2,tz);
1648
1649             fjx0             = _mm_add_ps(fjx0,tx);
1650             fjy0             = _mm_add_ps(fjy0,ty);
1651             fjz0             = _mm_add_ps(fjz0,tz);
1652
1653             /**************************
1654              * CALCULATE INTERACTIONS *
1655              **************************/
1656
1657             /* REACTION-FIELD ELECTROSTATICS */
1658             felec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_mul_ps(rinv21,rinvsq21),krf2));
1659
1660             fscal            = felec;
1661
1662             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1663
1664             /* Calculate temporary vectorial force */
1665             tx               = _mm_mul_ps(fscal,dx21);
1666             ty               = _mm_mul_ps(fscal,dy21);
1667             tz               = _mm_mul_ps(fscal,dz21);
1668
1669             /* Update vectorial force */
1670             fix2             = _mm_add_ps(fix2,tx);
1671             fiy2             = _mm_add_ps(fiy2,ty);
1672             fiz2             = _mm_add_ps(fiz2,tz);
1673
1674             fjx1             = _mm_add_ps(fjx1,tx);
1675             fjy1             = _mm_add_ps(fjy1,ty);
1676             fjz1             = _mm_add_ps(fjz1,tz);
1677
1678             /**************************
1679              * CALCULATE INTERACTIONS *
1680              **************************/
1681
1682             /* REACTION-FIELD ELECTROSTATICS */
1683             felec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_mul_ps(rinv22,rinvsq22),krf2));
1684
1685             fscal            = felec;
1686
1687             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1688
1689             /* Calculate temporary vectorial force */
1690             tx               = _mm_mul_ps(fscal,dx22);
1691             ty               = _mm_mul_ps(fscal,dy22);
1692             tz               = _mm_mul_ps(fscal,dz22);
1693
1694             /* Update vectorial force */
1695             fix2             = _mm_add_ps(fix2,tx);
1696             fiy2             = _mm_add_ps(fiy2,ty);
1697             fiz2             = _mm_add_ps(fiz2,tz);
1698
1699             fjx2             = _mm_add_ps(fjx2,tx);
1700             fjy2             = _mm_add_ps(fjy2,ty);
1701             fjz2             = _mm_add_ps(fjz2,tz);
1702
1703             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1704             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1705             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1706             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1707
1708             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1709                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1710
1711             /* Inner loop uses 243 flops */
1712         }
1713
1714         /* End of innermost loop */
1715
1716         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1717                                               f+i_coord_offset,fshift+i_shift_offset);
1718
1719         /* Increment number of inner iterations */
1720         inneriter                  += j_index_end - j_index_start;
1721
1722         /* Outer loop uses 18 flops */
1723     }
1724
1725     /* Increment number of outer iterations */
1726     outeriter        += nri;
1727
1728     /* Update outer/inner flops */
1729
1730     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W3W3_F,outeriter*18 + inneriter*243);
1731 }