src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_double/nb_kernel_ElecRFCut_VdwNone_GeomW4W4_sse4_1_double.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_double 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_double.h"
  48 #include "kernelutil_x86_sse4_1_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwNone_GeomW4W4_VF_sse4_1_double
  52  * Electrostatics interaction: ReactionField
  53  * VdW interaction:            None
  54  * Geometry:                   Water4-Water4
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecRFCut_VdwNone_GeomW4W4_VF_sse4_1_double
  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 refer to j loop unrolling done with SSE double precision, e.g. for the two 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;
  75     int              j_coord_offsetA,j_coord_offsetB;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  80     int              vdwioffset1;
  81     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  82     int              vdwioffset2;
  83     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  84     int              vdwioffset3;
  85     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  86     int              vdwjidx1A,vdwjidx1B;
  87     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  88     int              vdwjidx2A,vdwjidx2B;
  89     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  90     int              vdwjidx3A,vdwjidx3B;
  91     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  92     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  93     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  94     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  95     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  96     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  97     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  98     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  99     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 100     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 101     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 102     real             *charge;
 103     __m128d          dummy_mask,cutoff_mask;
 104     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 105     __m128d          one     = _mm_set1_pd(1.0);
 106     __m128d          two     = _mm_set1_pd(2.0);
 107     x                = xx[0];
 108     f                = ff[0];
 109
 110     nri              = nlist->nri;
 111     iinr             = nlist->iinr;
 112     jindex           = nlist->jindex;
 113     jjnr             = nlist->jjnr;
 114     shiftidx         = nlist->shift;
 115     gid              = nlist->gid;
 116     shiftvec         = fr->shift_vec[0];
 117     fshift           = fr->fshift[0];
 118     facel            = _mm_set1_pd(fr->epsfac);
 119     charge           = mdatoms->chargeA;
 120     krf              = _mm_set1_pd(fr->ic->k_rf);
 121     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 122     crf              = _mm_set1_pd(fr->ic->c_rf);
 123
 124     /* Setup water-specific parameters */
 125     inr              = nlist->iinr[0];
 126     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 127     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 128     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 129
 130     jq1              = _mm_set1_pd(charge[inr+1]);
 131     jq2              = _mm_set1_pd(charge[inr+2]);
 132     jq3              = _mm_set1_pd(charge[inr+3]);
 133     qq11             = _mm_mul_pd(iq1,jq1);
 134     qq12             = _mm_mul_pd(iq1,jq2);
 135     qq13             = _mm_mul_pd(iq1,jq3);
 136     qq21             = _mm_mul_pd(iq2,jq1);
 137     qq22             = _mm_mul_pd(iq2,jq2);
 138     qq23             = _mm_mul_pd(iq2,jq3);
 139     qq31             = _mm_mul_pd(iq3,jq1);
 140     qq32             = _mm_mul_pd(iq3,jq2);
 141     qq33             = _mm_mul_pd(iq3,jq3);
 142
 143     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 144     rcutoff_scalar   = fr->rcoulomb;
 145     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 146     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 147
 148     /* Avoid stupid compiler warnings */
 149     jnrA = jnrB = 0;
 150     j_coord_offsetA = 0;
 151     j_coord_offsetB = 0;
 152
 153     outeriter        = 0;
 154     inneriter        = 0;
 155
 156     /* Start outer loop over neighborlists */
 157     for(iidx=0; iidx<nri; iidx++)
 158     {
 159         /* Load shift vector for this list */
 160         i_shift_offset   = DIM*shiftidx[iidx];
 161
 162         /* Load limits for loop over neighbors */
 163         j_index_start    = jindex[iidx];
 164         j_index_end      = jindex[iidx+1];
 165
 166         /* Get outer coordinate index */
 167         inr              = iinr[iidx];
 168         i_coord_offset   = DIM*inr;
 169
 170         /* Load i particle coords and add shift vector */
 171         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 172                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 173
 174         fix1             = _mm_setzero_pd();
 175         fiy1             = _mm_setzero_pd();
 176         fiz1             = _mm_setzero_pd();
 177         fix2             = _mm_setzero_pd();
 178         fiy2             = _mm_setzero_pd();
 179         fiz2             = _mm_setzero_pd();
 180         fix3             = _mm_setzero_pd();
 181         fiy3             = _mm_setzero_pd();
 182         fiz3             = _mm_setzero_pd();
 183
 184         /* Reset potential sums */
 185         velecsum         = _mm_setzero_pd();
 186
 187         /* Start inner kernel loop */
 188         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 189         {
 190
 191             /* Get j neighbor index, and coordinate index */
 192             jnrA             = jjnr[jidx];
 193             jnrB             = jjnr[jidx+1];
 194             j_coord_offsetA  = DIM*jnrA;
 195             j_coord_offsetB  = DIM*jnrB;
 196
 197             /* load j atom coordinates */
 198             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 199                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 200
 201             /* Calculate displacement vector */
 202             dx11             = _mm_sub_pd(ix1,jx1);
 203             dy11             = _mm_sub_pd(iy1,jy1);
 204             dz11             = _mm_sub_pd(iz1,jz1);
 205             dx12             = _mm_sub_pd(ix1,jx2);
 206             dy12             = _mm_sub_pd(iy1,jy2);
 207             dz12             = _mm_sub_pd(iz1,jz2);
 208             dx13             = _mm_sub_pd(ix1,jx3);
 209             dy13             = _mm_sub_pd(iy1,jy3);
 210             dz13             = _mm_sub_pd(iz1,jz3);
 211             dx21             = _mm_sub_pd(ix2,jx1);
 212             dy21             = _mm_sub_pd(iy2,jy1);
 213             dz21             = _mm_sub_pd(iz2,jz1);
 214             dx22             = _mm_sub_pd(ix2,jx2);
 215             dy22             = _mm_sub_pd(iy2,jy2);
 216             dz22             = _mm_sub_pd(iz2,jz2);
 217             dx23             = _mm_sub_pd(ix2,jx3);
 218             dy23             = _mm_sub_pd(iy2,jy3);
 219             dz23             = _mm_sub_pd(iz2,jz3);
 220             dx31             = _mm_sub_pd(ix3,jx1);
 221             dy31             = _mm_sub_pd(iy3,jy1);
 222             dz31             = _mm_sub_pd(iz3,jz1);
 223             dx32             = _mm_sub_pd(ix3,jx2);
 224             dy32             = _mm_sub_pd(iy3,jy2);
 225             dz32             = _mm_sub_pd(iz3,jz2);
 226             dx33             = _mm_sub_pd(ix3,jx3);
 227             dy33             = _mm_sub_pd(iy3,jy3);
 228             dz33             = _mm_sub_pd(iz3,jz3);
 229
 230             /* Calculate squared distance and things based on it */
 231             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 232             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 233             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 234             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 235             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 236             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 237             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 238             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 239             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 240
 241             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 242             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 243             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 244             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 245             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 246             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 247             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 248             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 249             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 250
 251             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 252             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 253             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 254             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 255             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 256             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 257             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 258             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 259             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 260
 261             fjx1             = _mm_setzero_pd();
 262             fjy1             = _mm_setzero_pd();
 263             fjz1             = _mm_setzero_pd();
 264             fjx2             = _mm_setzero_pd();
 265             fjy2             = _mm_setzero_pd();
 266             fjz2             = _mm_setzero_pd();
 267             fjx3             = _mm_setzero_pd();
 268             fjy3             = _mm_setzero_pd();
 269             fjz3             = _mm_setzero_pd();
 270
 271             /**************************
 272              * CALCULATE INTERACTIONS *
 273              **************************/
 274
 275             if (gmx_mm_any_lt(rsq11,rcutoff2))
 276             {
 277
 278             /* REACTION-FIELD ELECTROSTATICS */
 279             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 280             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 281
 282             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 283
 284             /* Update potential sum for this i atom from the interaction with this j atom. */
 285             velec            = _mm_and_pd(velec,cutoff_mask);
 286             velecsum         = _mm_add_pd(velecsum,velec);
 287
 288             fscal            = felec;
 289
 290             fscal            = _mm_and_pd(fscal,cutoff_mask);
 291
 292             /* Calculate temporary vectorial force */
 293             tx               = _mm_mul_pd(fscal,dx11);
 294             ty               = _mm_mul_pd(fscal,dy11);
 295             tz               = _mm_mul_pd(fscal,dz11);
 296
 297             /* Update vectorial force */
 298             fix1             = _mm_add_pd(fix1,tx);
 299             fiy1             = _mm_add_pd(fiy1,ty);
 300             fiz1             = _mm_add_pd(fiz1,tz);
 301
 302             fjx1             = _mm_add_pd(fjx1,tx);
 303             fjy1             = _mm_add_pd(fjy1,ty);
 304             fjz1             = _mm_add_pd(fjz1,tz);
 305
 306             }
 307
 308             /**************************
 309              * CALCULATE INTERACTIONS *
 310              **************************/
 311
 312             if (gmx_mm_any_lt(rsq12,rcutoff2))
 313             {
 314
 315             /* REACTION-FIELD ELECTROSTATICS */
 316             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 317             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 318
 319             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 320
 321             /* Update potential sum for this i atom from the interaction with this j atom. */
 322             velec            = _mm_and_pd(velec,cutoff_mask);
 323             velecsum         = _mm_add_pd(velecsum,velec);
 324
 325             fscal            = felec;
 326
 327             fscal            = _mm_and_pd(fscal,cutoff_mask);
 328
 329             /* Calculate temporary vectorial force */
 330             tx               = _mm_mul_pd(fscal,dx12);
 331             ty               = _mm_mul_pd(fscal,dy12);
 332             tz               = _mm_mul_pd(fscal,dz12);
 333
 334             /* Update vectorial force */
 335             fix1             = _mm_add_pd(fix1,tx);
 336             fiy1             = _mm_add_pd(fiy1,ty);
 337             fiz1             = _mm_add_pd(fiz1,tz);
 338
 339             fjx2             = _mm_add_pd(fjx2,tx);
 340             fjy2             = _mm_add_pd(fjy2,ty);
 341             fjz2             = _mm_add_pd(fjz2,tz);
 342
 343             }
 344
 345             /**************************
 346              * CALCULATE INTERACTIONS *
 347              **************************/
 348
 349             if (gmx_mm_any_lt(rsq13,rcutoff2))
 350             {
 351
 352             /* REACTION-FIELD ELECTROSTATICS */
 353             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
 354             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
 355
 356             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 357
 358             /* Update potential sum for this i atom from the interaction with this j atom. */
 359             velec            = _mm_and_pd(velec,cutoff_mask);
 360             velecsum         = _mm_add_pd(velecsum,velec);
 361
 362             fscal            = felec;
 363
 364             fscal            = _mm_and_pd(fscal,cutoff_mask);
 365
 366             /* Calculate temporary vectorial force */
 367             tx               = _mm_mul_pd(fscal,dx13);
 368             ty               = _mm_mul_pd(fscal,dy13);
 369             tz               = _mm_mul_pd(fscal,dz13);
 370
 371             /* Update vectorial force */
 372             fix1             = _mm_add_pd(fix1,tx);
 373             fiy1             = _mm_add_pd(fiy1,ty);
 374             fiz1             = _mm_add_pd(fiz1,tz);
 375
 376             fjx3             = _mm_add_pd(fjx3,tx);
 377             fjy3             = _mm_add_pd(fjy3,ty);
 378             fjz3             = _mm_add_pd(fjz3,tz);
 379
 380             }
 381
 382             /**************************
 383              * CALCULATE INTERACTIONS *
 384              **************************/
 385
 386             if (gmx_mm_any_lt(rsq21,rcutoff2))
 387             {
 388
 389             /* REACTION-FIELD ELECTROSTATICS */
 390             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 391             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 392
 393             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 394
 395             /* Update potential sum for this i atom from the interaction with this j atom. */
 396             velec            = _mm_and_pd(velec,cutoff_mask);
 397             velecsum         = _mm_add_pd(velecsum,velec);
 398
 399             fscal            = felec;
 400
 401             fscal            = _mm_and_pd(fscal,cutoff_mask);
 402
 403             /* Calculate temporary vectorial force */
 404             tx               = _mm_mul_pd(fscal,dx21);
 405             ty               = _mm_mul_pd(fscal,dy21);
 406             tz               = _mm_mul_pd(fscal,dz21);
 407
 408             /* Update vectorial force */
 409             fix2             = _mm_add_pd(fix2,tx);
 410             fiy2             = _mm_add_pd(fiy2,ty);
 411             fiz2             = _mm_add_pd(fiz2,tz);
 412
 413             fjx1             = _mm_add_pd(fjx1,tx);
 414             fjy1             = _mm_add_pd(fjy1,ty);
 415             fjz1             = _mm_add_pd(fjz1,tz);
 416
 417             }
 418
 419             /**************************
 420              * CALCULATE INTERACTIONS *
 421              **************************/
 422
 423             if (gmx_mm_any_lt(rsq22,rcutoff2))
 424             {
 425
 426             /* REACTION-FIELD ELECTROSTATICS */
 427             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 428             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 429
 430             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 431
 432             /* Update potential sum for this i atom from the interaction with this j atom. */
 433             velec            = _mm_and_pd(velec,cutoff_mask);
 434             velecsum         = _mm_add_pd(velecsum,velec);
 435
 436             fscal            = felec;
 437
 438             fscal            = _mm_and_pd(fscal,cutoff_mask);
 439
 440             /* Calculate temporary vectorial force */
 441             tx               = _mm_mul_pd(fscal,dx22);
 442             ty               = _mm_mul_pd(fscal,dy22);
 443             tz               = _mm_mul_pd(fscal,dz22);
 444
 445             /* Update vectorial force */
 446             fix2             = _mm_add_pd(fix2,tx);
 447             fiy2             = _mm_add_pd(fiy2,ty);
 448             fiz2             = _mm_add_pd(fiz2,tz);
 449
 450             fjx2             = _mm_add_pd(fjx2,tx);
 451             fjy2             = _mm_add_pd(fjy2,ty);
 452             fjz2             = _mm_add_pd(fjz2,tz);
 453
 454             }
 455
 456             /**************************
 457              * CALCULATE INTERACTIONS *
 458              **************************/
 459
 460             if (gmx_mm_any_lt(rsq23,rcutoff2))
 461             {
 462
 463             /* REACTION-FIELD ELECTROSTATICS */
 464             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
 465             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
 466
 467             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 468
 469             /* Update potential sum for this i atom from the interaction with this j atom. */
 470             velec            = _mm_and_pd(velec,cutoff_mask);
 471             velecsum         = _mm_add_pd(velecsum,velec);
 472
 473             fscal            = felec;
 474
 475             fscal            = _mm_and_pd(fscal,cutoff_mask);
 476
 477             /* Calculate temporary vectorial force */
 478             tx               = _mm_mul_pd(fscal,dx23);
 479             ty               = _mm_mul_pd(fscal,dy23);
 480             tz               = _mm_mul_pd(fscal,dz23);
 481
 482             /* Update vectorial force */
 483             fix2             = _mm_add_pd(fix2,tx);
 484             fiy2             = _mm_add_pd(fiy2,ty);
 485             fiz2             = _mm_add_pd(fiz2,tz);
 486
 487             fjx3             = _mm_add_pd(fjx3,tx);
 488             fjy3             = _mm_add_pd(fjy3,ty);
 489             fjz3             = _mm_add_pd(fjz3,tz);
 490
 491             }
 492
 493             /**************************
 494              * CALCULATE INTERACTIONS *
 495              **************************/
 496
 497             if (gmx_mm_any_lt(rsq31,rcutoff2))
 498             {
 499
 500             /* REACTION-FIELD ELECTROSTATICS */
 501             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
 502             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
 503
 504             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
 505
 506             /* Update potential sum for this i atom from the interaction with this j atom. */
 507             velec            = _mm_and_pd(velec,cutoff_mask);
 508             velecsum         = _mm_add_pd(velecsum,velec);
 509
 510             fscal            = felec;
 511
 512             fscal            = _mm_and_pd(fscal,cutoff_mask);
 513
 514             /* Calculate temporary vectorial force */
 515             tx               = _mm_mul_pd(fscal,dx31);
 516             ty               = _mm_mul_pd(fscal,dy31);
 517             tz               = _mm_mul_pd(fscal,dz31);
 518
 519             /* Update vectorial force */
 520             fix3             = _mm_add_pd(fix3,tx);
 521             fiy3             = _mm_add_pd(fiy3,ty);
 522             fiz3             = _mm_add_pd(fiz3,tz);
 523
 524             fjx1             = _mm_add_pd(fjx1,tx);
 525             fjy1             = _mm_add_pd(fjy1,ty);
 526             fjz1             = _mm_add_pd(fjz1,tz);
 527
 528             }
 529
 530             /**************************
 531              * CALCULATE INTERACTIONS *
 532              **************************/
 533
 534             if (gmx_mm_any_lt(rsq32,rcutoff2))
 535             {
 536
 537             /* REACTION-FIELD ELECTROSTATICS */
 538             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
 539             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
 540
 541             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
 542
 543             /* Update potential sum for this i atom from the interaction with this j atom. */
 544             velec            = _mm_and_pd(velec,cutoff_mask);
 545             velecsum         = _mm_add_pd(velecsum,velec);
 546
 547             fscal            = felec;
 548
 549             fscal            = _mm_and_pd(fscal,cutoff_mask);
 550
 551             /* Calculate temporary vectorial force */
 552             tx               = _mm_mul_pd(fscal,dx32);
 553             ty               = _mm_mul_pd(fscal,dy32);
 554             tz               = _mm_mul_pd(fscal,dz32);
 555
 556             /* Update vectorial force */
 557             fix3             = _mm_add_pd(fix3,tx);
 558             fiy3             = _mm_add_pd(fiy3,ty);
 559             fiz3             = _mm_add_pd(fiz3,tz);
 560
 561             fjx2             = _mm_add_pd(fjx2,tx);
 562             fjy2             = _mm_add_pd(fjy2,ty);
 563             fjz2             = _mm_add_pd(fjz2,tz);
 564
 565             }
 566
 567             /**************************
 568              * CALCULATE INTERACTIONS *
 569              **************************/
 570
 571             if (gmx_mm_any_lt(rsq33,rcutoff2))
 572             {
 573
 574             /* REACTION-FIELD ELECTROSTATICS */
 575             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
 576             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
 577
 578             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
 579
 580             /* Update potential sum for this i atom from the interaction with this j atom. */
 581             velec            = _mm_and_pd(velec,cutoff_mask);
 582             velecsum         = _mm_add_pd(velecsum,velec);
 583
 584             fscal            = felec;
 585
 586             fscal            = _mm_and_pd(fscal,cutoff_mask);
 587
 588             /* Calculate temporary vectorial force */
 589             tx               = _mm_mul_pd(fscal,dx33);
 590             ty               = _mm_mul_pd(fscal,dy33);
 591             tz               = _mm_mul_pd(fscal,dz33);
 592
 593             /* Update vectorial force */
 594             fix3             = _mm_add_pd(fix3,tx);
 595             fiy3             = _mm_add_pd(fiy3,ty);
 596             fiz3             = _mm_add_pd(fiz3,tz);
 597
 598             fjx3             = _mm_add_pd(fjx3,tx);
 599             fjy3             = _mm_add_pd(fjy3,ty);
 600             fjz3             = _mm_add_pd(fjz3,tz);
 601
 602             }
 603
 604             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 605
 606             /* Inner loop uses 324 flops */
 607         }
 608
 609         if(jidx<j_index_end)
 610         {
 611
 612             jnrA             = jjnr[jidx];
 613             j_coord_offsetA  = DIM*jnrA;
 614
 615             /* load j atom coordinates */
 616             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
 617                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 618
 619             /* Calculate displacement vector */
 620             dx11             = _mm_sub_pd(ix1,jx1);
 621             dy11             = _mm_sub_pd(iy1,jy1);
 622             dz11             = _mm_sub_pd(iz1,jz1);
 623             dx12             = _mm_sub_pd(ix1,jx2);
 624             dy12             = _mm_sub_pd(iy1,jy2);
 625             dz12             = _mm_sub_pd(iz1,jz2);
 626             dx13             = _mm_sub_pd(ix1,jx3);
 627             dy13             = _mm_sub_pd(iy1,jy3);
 628             dz13             = _mm_sub_pd(iz1,jz3);
 629             dx21             = _mm_sub_pd(ix2,jx1);
 630             dy21             = _mm_sub_pd(iy2,jy1);
 631             dz21             = _mm_sub_pd(iz2,jz1);
 632             dx22             = _mm_sub_pd(ix2,jx2);
 633             dy22             = _mm_sub_pd(iy2,jy2);
 634             dz22             = _mm_sub_pd(iz2,jz2);
 635             dx23             = _mm_sub_pd(ix2,jx3);
 636             dy23             = _mm_sub_pd(iy2,jy3);
 637             dz23             = _mm_sub_pd(iz2,jz3);
 638             dx31             = _mm_sub_pd(ix3,jx1);
 639             dy31             = _mm_sub_pd(iy3,jy1);
 640             dz31             = _mm_sub_pd(iz3,jz1);
 641             dx32             = _mm_sub_pd(ix3,jx2);
 642             dy32             = _mm_sub_pd(iy3,jy2);
 643             dz32             = _mm_sub_pd(iz3,jz2);
 644             dx33             = _mm_sub_pd(ix3,jx3);
 645             dy33             = _mm_sub_pd(iy3,jy3);
 646             dz33             = _mm_sub_pd(iz3,jz3);
 647
 648             /* Calculate squared distance and things based on it */
 649             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 650             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 651             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 652             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 653             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 654             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 655             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 656             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 657             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 658
 659             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 660             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 661             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 662             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 663             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 664             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 665             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 666             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 667             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 668
 669             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 670             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 671             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 672             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 673             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 674             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 675             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 676             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 677             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 678
 679             fjx1             = _mm_setzero_pd();
 680             fjy1             = _mm_setzero_pd();
 681             fjz1             = _mm_setzero_pd();
 682             fjx2             = _mm_setzero_pd();
 683             fjy2             = _mm_setzero_pd();
 684             fjz2             = _mm_setzero_pd();
 685             fjx3             = _mm_setzero_pd();
 686             fjy3             = _mm_setzero_pd();
 687             fjz3             = _mm_setzero_pd();
 688
 689             /**************************
 690              * CALCULATE INTERACTIONS *
 691              **************************/
 692
 693             if (gmx_mm_any_lt(rsq11,rcutoff2))
 694             {
 695
 696             /* REACTION-FIELD ELECTROSTATICS */
 697             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 698             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 699
 700             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 701
 702             /* Update potential sum for this i atom from the interaction with this j atom. */
 703             velec            = _mm_and_pd(velec,cutoff_mask);
 704             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 705             velecsum         = _mm_add_pd(velecsum,velec);
 706
 707             fscal            = felec;
 708
 709             fscal            = _mm_and_pd(fscal,cutoff_mask);
 710
 711             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 712
 713             /* Calculate temporary vectorial force */
 714             tx               = _mm_mul_pd(fscal,dx11);
 715             ty               = _mm_mul_pd(fscal,dy11);
 716             tz               = _mm_mul_pd(fscal,dz11);
 717
 718             /* Update vectorial force */
 719             fix1             = _mm_add_pd(fix1,tx);
 720             fiy1             = _mm_add_pd(fiy1,ty);
 721             fiz1             = _mm_add_pd(fiz1,tz);
 722
 723             fjx1             = _mm_add_pd(fjx1,tx);
 724             fjy1             = _mm_add_pd(fjy1,ty);
 725             fjz1             = _mm_add_pd(fjz1,tz);
 726
 727             }
 728
 729             /**************************
 730              * CALCULATE INTERACTIONS *
 731              **************************/
 732
 733             if (gmx_mm_any_lt(rsq12,rcutoff2))
 734             {
 735
 736             /* REACTION-FIELD ELECTROSTATICS */
 737             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 738             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 739
 740             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 741
 742             /* Update potential sum for this i atom from the interaction with this j atom. */
 743             velec            = _mm_and_pd(velec,cutoff_mask);
 744             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 745             velecsum         = _mm_add_pd(velecsum,velec);
 746
 747             fscal            = felec;
 748
 749             fscal            = _mm_and_pd(fscal,cutoff_mask);
 750
 751             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 752
 753             /* Calculate temporary vectorial force */
 754             tx               = _mm_mul_pd(fscal,dx12);
 755             ty               = _mm_mul_pd(fscal,dy12);
 756             tz               = _mm_mul_pd(fscal,dz12);
 757
 758             /* Update vectorial force */
 759             fix1             = _mm_add_pd(fix1,tx);
 760             fiy1             = _mm_add_pd(fiy1,ty);
 761             fiz1             = _mm_add_pd(fiz1,tz);
 762
 763             fjx2             = _mm_add_pd(fjx2,tx);
 764             fjy2             = _mm_add_pd(fjy2,ty);
 765             fjz2             = _mm_add_pd(fjz2,tz);
 766
 767             }
 768
 769             /**************************
 770              * CALCULATE INTERACTIONS *
 771              **************************/
 772
 773             if (gmx_mm_any_lt(rsq13,rcutoff2))
 774             {
 775
 776             /* REACTION-FIELD ELECTROSTATICS */
 777             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
 778             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
 779
 780             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 781
 782             /* Update potential sum for this i atom from the interaction with this j atom. */
 783             velec            = _mm_and_pd(velec,cutoff_mask);
 784             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 785             velecsum         = _mm_add_pd(velecsum,velec);
 786
 787             fscal            = felec;
 788
 789             fscal            = _mm_and_pd(fscal,cutoff_mask);
 790
 791             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 792
 793             /* Calculate temporary vectorial force */
 794             tx               = _mm_mul_pd(fscal,dx13);
 795             ty               = _mm_mul_pd(fscal,dy13);
 796             tz               = _mm_mul_pd(fscal,dz13);
 797
 798             /* Update vectorial force */
 799             fix1             = _mm_add_pd(fix1,tx);
 800             fiy1             = _mm_add_pd(fiy1,ty);
 801             fiz1             = _mm_add_pd(fiz1,tz);
 802
 803             fjx3             = _mm_add_pd(fjx3,tx);
 804             fjy3             = _mm_add_pd(fjy3,ty);
 805             fjz3             = _mm_add_pd(fjz3,tz);
 806
 807             }
 808
 809             /**************************
 810              * CALCULATE INTERACTIONS *
 811              **************************/
 812
 813             if (gmx_mm_any_lt(rsq21,rcutoff2))
 814             {
 815
 816             /* REACTION-FIELD ELECTROSTATICS */
 817             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 818             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 819
 820             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 821
 822             /* Update potential sum for this i atom from the interaction with this j atom. */
 823             velec            = _mm_and_pd(velec,cutoff_mask);
 824             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 825             velecsum         = _mm_add_pd(velecsum,velec);
 826
 827             fscal            = felec;
 828
 829             fscal            = _mm_and_pd(fscal,cutoff_mask);
 830
 831             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 832
 833             /* Calculate temporary vectorial force */
 834             tx               = _mm_mul_pd(fscal,dx21);
 835             ty               = _mm_mul_pd(fscal,dy21);
 836             tz               = _mm_mul_pd(fscal,dz21);
 837
 838             /* Update vectorial force */
 839             fix2             = _mm_add_pd(fix2,tx);
 840             fiy2             = _mm_add_pd(fiy2,ty);
 841             fiz2             = _mm_add_pd(fiz2,tz);
 842
 843             fjx1             = _mm_add_pd(fjx1,tx);
 844             fjy1             = _mm_add_pd(fjy1,ty);
 845             fjz1             = _mm_add_pd(fjz1,tz);
 846
 847             }
 848
 849             /**************************
 850              * CALCULATE INTERACTIONS *
 851              **************************/
 852
 853             if (gmx_mm_any_lt(rsq22,rcutoff2))
 854             {
 855
 856             /* REACTION-FIELD ELECTROSTATICS */
 857             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 858             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 859
 860             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 861
 862             /* Update potential sum for this i atom from the interaction with this j atom. */
 863             velec            = _mm_and_pd(velec,cutoff_mask);
 864             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 865             velecsum         = _mm_add_pd(velecsum,velec);
 866
 867             fscal            = felec;
 868
 869             fscal            = _mm_and_pd(fscal,cutoff_mask);
 870
 871             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 872
 873             /* Calculate temporary vectorial force */
 874             tx               = _mm_mul_pd(fscal,dx22);
 875             ty               = _mm_mul_pd(fscal,dy22);
 876             tz               = _mm_mul_pd(fscal,dz22);
 877
 878             /* Update vectorial force */
 879             fix2             = _mm_add_pd(fix2,tx);
 880             fiy2             = _mm_add_pd(fiy2,ty);
 881             fiz2             = _mm_add_pd(fiz2,tz);
 882
 883             fjx2             = _mm_add_pd(fjx2,tx);
 884             fjy2             = _mm_add_pd(fjy2,ty);
 885             fjz2             = _mm_add_pd(fjz2,tz);
 886
 887             }
 888
 889             /**************************
 890              * CALCULATE INTERACTIONS *
 891              **************************/
 892
 893             if (gmx_mm_any_lt(rsq23,rcutoff2))
 894             {
 895
 896             /* REACTION-FIELD ELECTROSTATICS */
 897             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
 898             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
 899
 900             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 901
 902             /* Update potential sum for this i atom from the interaction with this j atom. */
 903             velec            = _mm_and_pd(velec,cutoff_mask);
 904             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 905             velecsum         = _mm_add_pd(velecsum,velec);
 906
 907             fscal            = felec;
 908
 909             fscal            = _mm_and_pd(fscal,cutoff_mask);
 910
 911             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 912
 913             /* Calculate temporary vectorial force */
 914             tx               = _mm_mul_pd(fscal,dx23);
 915             ty               = _mm_mul_pd(fscal,dy23);
 916             tz               = _mm_mul_pd(fscal,dz23);
 917
 918             /* Update vectorial force */
 919             fix2             = _mm_add_pd(fix2,tx);
 920             fiy2             = _mm_add_pd(fiy2,ty);
 921             fiz2             = _mm_add_pd(fiz2,tz);
 922
 923             fjx3             = _mm_add_pd(fjx3,tx);
 924             fjy3             = _mm_add_pd(fjy3,ty);
 925             fjz3             = _mm_add_pd(fjz3,tz);
 926
 927             }
 928
 929             /**************************
 930              * CALCULATE INTERACTIONS *
 931              **************************/
 932
 933             if (gmx_mm_any_lt(rsq31,rcutoff2))
 934             {
 935
 936             /* REACTION-FIELD ELECTROSTATICS */
 937             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
 938             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
 939
 940             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
 941
 942             /* Update potential sum for this i atom from the interaction with this j atom. */
 943             velec            = _mm_and_pd(velec,cutoff_mask);
 944             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 945             velecsum         = _mm_add_pd(velecsum,velec);
 946
 947             fscal            = felec;
 948
 949             fscal            = _mm_and_pd(fscal,cutoff_mask);
 950
 951             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 952
 953             /* Calculate temporary vectorial force */
 954             tx               = _mm_mul_pd(fscal,dx31);
 955             ty               = _mm_mul_pd(fscal,dy31);
 956             tz               = _mm_mul_pd(fscal,dz31);
 957
 958             /* Update vectorial force */
 959             fix3             = _mm_add_pd(fix3,tx);
 960             fiy3             = _mm_add_pd(fiy3,ty);
 961             fiz3             = _mm_add_pd(fiz3,tz);
 962
 963             fjx1             = _mm_add_pd(fjx1,tx);
 964             fjy1             = _mm_add_pd(fjy1,ty);
 965             fjz1             = _mm_add_pd(fjz1,tz);
 966
 967             }
 968
 969             /**************************
 970              * CALCULATE INTERACTIONS *
 971              **************************/
 972
 973             if (gmx_mm_any_lt(rsq32,rcutoff2))
 974             {
 975
 976             /* REACTION-FIELD ELECTROSTATICS */
 977             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
 978             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
 979
 980             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
 981
 982             /* Update potential sum for this i atom from the interaction with this j atom. */
 983             velec            = _mm_and_pd(velec,cutoff_mask);
 984             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 985             velecsum         = _mm_add_pd(velecsum,velec);
 986
 987             fscal            = felec;
 988
 989             fscal            = _mm_and_pd(fscal,cutoff_mask);
 990
 991             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 992
 993             /* Calculate temporary vectorial force */
 994             tx               = _mm_mul_pd(fscal,dx32);
 995             ty               = _mm_mul_pd(fscal,dy32);
 996             tz               = _mm_mul_pd(fscal,dz32);
 997
 998             /* Update vectorial force */
 999             fix3             = _mm_add_pd(fix3,tx);
1000             fiy3             = _mm_add_pd(fiy3,ty);
1001             fiz3             = _mm_add_pd(fiz3,tz);
1002
1003             fjx2             = _mm_add_pd(fjx2,tx);
1004             fjy2             = _mm_add_pd(fjy2,ty);
1005             fjz2             = _mm_add_pd(fjz2,tz);
1006
1007             }
1008
1009             /**************************
1010              * CALCULATE INTERACTIONS *
1011              **************************/
1012
1013             if (gmx_mm_any_lt(rsq33,rcutoff2))
1014             {
1015
1016             /* REACTION-FIELD ELECTROSTATICS */
1017             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
1018             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1019
1020             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1021
1022             /* Update potential sum for this i atom from the interaction with this j atom. */
1023             velec            = _mm_and_pd(velec,cutoff_mask);
1024             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1025             velecsum         = _mm_add_pd(velecsum,velec);
1026
1027             fscal            = felec;
1028
1029             fscal            = _mm_and_pd(fscal,cutoff_mask);
1030
1031             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1032
1033             /* Calculate temporary vectorial force */
1034             tx               = _mm_mul_pd(fscal,dx33);
1035             ty               = _mm_mul_pd(fscal,dy33);
1036             tz               = _mm_mul_pd(fscal,dz33);
1037
1038             /* Update vectorial force */
1039             fix3             = _mm_add_pd(fix3,tx);
1040             fiy3             = _mm_add_pd(fiy3,ty);
1041             fiz3             = _mm_add_pd(fiz3,tz);
1042
1043             fjx3             = _mm_add_pd(fjx3,tx);
1044             fjy3             = _mm_add_pd(fjy3,ty);
1045             fjz3             = _mm_add_pd(fjz3,tz);
1046
1047             }
1048
1049             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1050
1051             /* Inner loop uses 324 flops */
1052         }
1053
1054         /* End of innermost loop */
1055
1056         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1057                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1058
1059         ggid                        = gid[iidx];
1060         /* Update potential energies */
1061         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1062
1063         /* Increment number of inner iterations */
1064         inneriter                  += j_index_end - j_index_start;
1065
1066         /* Outer loop uses 19 flops */
1067     }
1068
1069     /* Increment number of outer iterations */
1070     outeriter        += nri;
1071
1072     /* Update outer/inner flops */
1073
1074     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*324);
1075 }
1076 /*
1077  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwNone_GeomW4W4_F_sse4_1_double
1078  * Electrostatics interaction: ReactionField
1079  * VdW interaction:            None
1080  * Geometry:                   Water4-Water4
1081  * Calculate force/pot:        Force
1082  */
1083 void
1084 nb_kernel_ElecRFCut_VdwNone_GeomW4W4_F_sse4_1_double
1085                     (t_nblist                    * gmx_restrict       nlist,
1086                      rvec                        * gmx_restrict          xx,
1087                      rvec                        * gmx_restrict          ff,
1088                      t_forcerec                  * gmx_restrict          fr,
1089                      t_mdatoms                   * gmx_restrict     mdatoms,
1090                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1091                      t_nrnb                      * gmx_restrict        nrnb)
1092 {
1093     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1094      * just 0 for non-waters.
1095      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1096      * jnr indices corresponding to data put in the four positions in the SIMD register.
1097      */
1098     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1099     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1100     int              jnrA,jnrB;
1101     int              j_coord_offsetA,j_coord_offsetB;
1102     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1103     real             rcutoff_scalar;
1104     real             *shiftvec,*fshift,*x,*f;
1105     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1106     int              vdwioffset1;
1107     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1108     int              vdwioffset2;
1109     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1110     int              vdwioffset3;
1111     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1112     int              vdwjidx1A,vdwjidx1B;
1113     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1114     int              vdwjidx2A,vdwjidx2B;
1115     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1116     int              vdwjidx3A,vdwjidx3B;
1117     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1118     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1119     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1120     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1121     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1122     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1123     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1124     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1125     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1126     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1127     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1128     real             *charge;
1129     __m128d          dummy_mask,cutoff_mask;
1130     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1131     __m128d          one     = _mm_set1_pd(1.0);
1132     __m128d          two     = _mm_set1_pd(2.0);
1133     x                = xx[0];
1134     f                = ff[0];
1135
1136     nri              = nlist->nri;
1137     iinr             = nlist->iinr;
1138     jindex           = nlist->jindex;
1139     jjnr             = nlist->jjnr;
1140     shiftidx         = nlist->shift;
1141     gid              = nlist->gid;
1142     shiftvec         = fr->shift_vec[0];
1143     fshift           = fr->fshift[0];
1144     facel            = _mm_set1_pd(fr->epsfac);
1145     charge           = mdatoms->chargeA;
1146     krf              = _mm_set1_pd(fr->ic->k_rf);
1147     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1148     crf              = _mm_set1_pd(fr->ic->c_rf);
1149
1150     /* Setup water-specific parameters */
1151     inr              = nlist->iinr[0];
1152     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1153     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1154     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1155
1156     jq1              = _mm_set1_pd(charge[inr+1]);
1157     jq2              = _mm_set1_pd(charge[inr+2]);
1158     jq3              = _mm_set1_pd(charge[inr+3]);
1159     qq11             = _mm_mul_pd(iq1,jq1);
1160     qq12             = _mm_mul_pd(iq1,jq2);
1161     qq13             = _mm_mul_pd(iq1,jq3);
1162     qq21             = _mm_mul_pd(iq2,jq1);
1163     qq22             = _mm_mul_pd(iq2,jq2);
1164     qq23             = _mm_mul_pd(iq2,jq3);
1165     qq31             = _mm_mul_pd(iq3,jq1);
1166     qq32             = _mm_mul_pd(iq3,jq2);
1167     qq33             = _mm_mul_pd(iq3,jq3);
1168
1169     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1170     rcutoff_scalar   = fr->rcoulomb;
1171     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1172     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1173
1174     /* Avoid stupid compiler warnings */
1175     jnrA = jnrB = 0;
1176     j_coord_offsetA = 0;
1177     j_coord_offsetB = 0;
1178
1179     outeriter        = 0;
1180     inneriter        = 0;
1181
1182     /* Start outer loop over neighborlists */
1183     for(iidx=0; iidx<nri; iidx++)
1184     {
1185         /* Load shift vector for this list */
1186         i_shift_offset   = DIM*shiftidx[iidx];
1187
1188         /* Load limits for loop over neighbors */
1189         j_index_start    = jindex[iidx];
1190         j_index_end      = jindex[iidx+1];
1191
1192         /* Get outer coordinate index */
1193         inr              = iinr[iidx];
1194         i_coord_offset   = DIM*inr;
1195
1196         /* Load i particle coords and add shift vector */
1197         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
1198                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1199
1200         fix1             = _mm_setzero_pd();
1201         fiy1             = _mm_setzero_pd();
1202         fiz1             = _mm_setzero_pd();
1203         fix2             = _mm_setzero_pd();
1204         fiy2             = _mm_setzero_pd();
1205         fiz2             = _mm_setzero_pd();
1206         fix3             = _mm_setzero_pd();
1207         fiy3             = _mm_setzero_pd();
1208         fiz3             = _mm_setzero_pd();
1209
1210         /* Start inner kernel loop */
1211         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1212         {
1213
1214             /* Get j neighbor index, and coordinate index */
1215             jnrA             = jjnr[jidx];
1216             jnrB             = jjnr[jidx+1];
1217             j_coord_offsetA  = DIM*jnrA;
1218             j_coord_offsetB  = DIM*jnrB;
1219
1220             /* load j atom coordinates */
1221             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1222                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1223
1224             /* Calculate displacement vector */
1225             dx11             = _mm_sub_pd(ix1,jx1);
1226             dy11             = _mm_sub_pd(iy1,jy1);
1227             dz11             = _mm_sub_pd(iz1,jz1);
1228             dx12             = _mm_sub_pd(ix1,jx2);
1229             dy12             = _mm_sub_pd(iy1,jy2);
1230             dz12             = _mm_sub_pd(iz1,jz2);
1231             dx13             = _mm_sub_pd(ix1,jx3);
1232             dy13             = _mm_sub_pd(iy1,jy3);
1233             dz13             = _mm_sub_pd(iz1,jz3);
1234             dx21             = _mm_sub_pd(ix2,jx1);
1235             dy21             = _mm_sub_pd(iy2,jy1);
1236             dz21             = _mm_sub_pd(iz2,jz1);
1237             dx22             = _mm_sub_pd(ix2,jx2);
1238             dy22             = _mm_sub_pd(iy2,jy2);
1239             dz22             = _mm_sub_pd(iz2,jz2);
1240             dx23             = _mm_sub_pd(ix2,jx3);
1241             dy23             = _mm_sub_pd(iy2,jy3);
1242             dz23             = _mm_sub_pd(iz2,jz3);
1243             dx31             = _mm_sub_pd(ix3,jx1);
1244             dy31             = _mm_sub_pd(iy3,jy1);
1245             dz31             = _mm_sub_pd(iz3,jz1);
1246             dx32             = _mm_sub_pd(ix3,jx2);
1247             dy32             = _mm_sub_pd(iy3,jy2);
1248             dz32             = _mm_sub_pd(iz3,jz2);
1249             dx33             = _mm_sub_pd(ix3,jx3);
1250             dy33             = _mm_sub_pd(iy3,jy3);
1251             dz33             = _mm_sub_pd(iz3,jz3);
1252
1253             /* Calculate squared distance and things based on it */
1254             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1255             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1256             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1257             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1258             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1259             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1260             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1261             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1262             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1263
1264             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1265             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1266             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1267             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1268             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1269             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1270             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1271             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1272             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1273
1274             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1275             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1276             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1277             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1278             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1279             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1280             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1281             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1282             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1283
1284             fjx1             = _mm_setzero_pd();
1285             fjy1             = _mm_setzero_pd();
1286             fjz1             = _mm_setzero_pd();
1287             fjx2             = _mm_setzero_pd();
1288             fjy2             = _mm_setzero_pd();
1289             fjz2             = _mm_setzero_pd();
1290             fjx3             = _mm_setzero_pd();
1291             fjy3             = _mm_setzero_pd();
1292             fjz3             = _mm_setzero_pd();
1293
1294             /**************************
1295              * CALCULATE INTERACTIONS *
1296              **************************/
1297
1298             if (gmx_mm_any_lt(rsq11,rcutoff2))
1299             {
1300
1301             /* REACTION-FIELD ELECTROSTATICS */
1302             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1303
1304             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1305
1306             fscal            = felec;
1307
1308             fscal            = _mm_and_pd(fscal,cutoff_mask);
1309
1310             /* Calculate temporary vectorial force */
1311             tx               = _mm_mul_pd(fscal,dx11);
1312             ty               = _mm_mul_pd(fscal,dy11);
1313             tz               = _mm_mul_pd(fscal,dz11);
1314
1315             /* Update vectorial force */
1316             fix1             = _mm_add_pd(fix1,tx);
1317             fiy1             = _mm_add_pd(fiy1,ty);
1318             fiz1             = _mm_add_pd(fiz1,tz);
1319
1320             fjx1             = _mm_add_pd(fjx1,tx);
1321             fjy1             = _mm_add_pd(fjy1,ty);
1322             fjz1             = _mm_add_pd(fjz1,tz);
1323
1324             }
1325
1326             /**************************
1327              * CALCULATE INTERACTIONS *
1328              **************************/
1329
1330             if (gmx_mm_any_lt(rsq12,rcutoff2))
1331             {
1332
1333             /* REACTION-FIELD ELECTROSTATICS */
1334             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1335
1336             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1337
1338             fscal            = felec;
1339
1340             fscal            = _mm_and_pd(fscal,cutoff_mask);
1341
1342             /* Calculate temporary vectorial force */
1343             tx               = _mm_mul_pd(fscal,dx12);
1344             ty               = _mm_mul_pd(fscal,dy12);
1345             tz               = _mm_mul_pd(fscal,dz12);
1346
1347             /* Update vectorial force */
1348             fix1             = _mm_add_pd(fix1,tx);
1349             fiy1             = _mm_add_pd(fiy1,ty);
1350             fiz1             = _mm_add_pd(fiz1,tz);
1351
1352             fjx2             = _mm_add_pd(fjx2,tx);
1353             fjy2             = _mm_add_pd(fjy2,ty);
1354             fjz2             = _mm_add_pd(fjz2,tz);
1355
1356             }
1357
1358             /**************************
1359              * CALCULATE INTERACTIONS *
1360              **************************/
1361
1362             if (gmx_mm_any_lt(rsq13,rcutoff2))
1363             {
1364
1365             /* REACTION-FIELD ELECTROSTATICS */
1366             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
1367
1368             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1369
1370             fscal            = felec;
1371
1372             fscal            = _mm_and_pd(fscal,cutoff_mask);
1373
1374             /* Calculate temporary vectorial force */
1375             tx               = _mm_mul_pd(fscal,dx13);
1376             ty               = _mm_mul_pd(fscal,dy13);
1377             tz               = _mm_mul_pd(fscal,dz13);
1378
1379             /* Update vectorial force */
1380             fix1             = _mm_add_pd(fix1,tx);
1381             fiy1             = _mm_add_pd(fiy1,ty);
1382             fiz1             = _mm_add_pd(fiz1,tz);
1383
1384             fjx3             = _mm_add_pd(fjx3,tx);
1385             fjy3             = _mm_add_pd(fjy3,ty);
1386             fjz3             = _mm_add_pd(fjz3,tz);
1387
1388             }
1389
1390             /**************************
1391              * CALCULATE INTERACTIONS *
1392              **************************/
1393
1394             if (gmx_mm_any_lt(rsq21,rcutoff2))
1395             {
1396
1397             /* REACTION-FIELD ELECTROSTATICS */
1398             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1399
1400             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1401
1402             fscal            = felec;
1403
1404             fscal            = _mm_and_pd(fscal,cutoff_mask);
1405
1406             /* Calculate temporary vectorial force */
1407             tx               = _mm_mul_pd(fscal,dx21);
1408             ty               = _mm_mul_pd(fscal,dy21);
1409             tz               = _mm_mul_pd(fscal,dz21);
1410
1411             /* Update vectorial force */
1412             fix2             = _mm_add_pd(fix2,tx);
1413             fiy2             = _mm_add_pd(fiy2,ty);
1414             fiz2             = _mm_add_pd(fiz2,tz);
1415
1416             fjx1             = _mm_add_pd(fjx1,tx);
1417             fjy1             = _mm_add_pd(fjy1,ty);
1418             fjz1             = _mm_add_pd(fjz1,tz);
1419
1420             }
1421
1422             /**************************
1423              * CALCULATE INTERACTIONS *
1424              **************************/
1425
1426             if (gmx_mm_any_lt(rsq22,rcutoff2))
1427             {
1428
1429             /* REACTION-FIELD ELECTROSTATICS */
1430             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1431
1432             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1433
1434             fscal            = felec;
1435
1436             fscal            = _mm_and_pd(fscal,cutoff_mask);
1437
1438             /* Calculate temporary vectorial force */
1439             tx               = _mm_mul_pd(fscal,dx22);
1440             ty               = _mm_mul_pd(fscal,dy22);
1441             tz               = _mm_mul_pd(fscal,dz22);
1442
1443             /* Update vectorial force */
1444             fix2             = _mm_add_pd(fix2,tx);
1445             fiy2             = _mm_add_pd(fiy2,ty);
1446             fiz2             = _mm_add_pd(fiz2,tz);
1447
1448             fjx2             = _mm_add_pd(fjx2,tx);
1449             fjy2             = _mm_add_pd(fjy2,ty);
1450             fjz2             = _mm_add_pd(fjz2,tz);
1451
1452             }
1453
1454             /**************************
1455              * CALCULATE INTERACTIONS *
1456              **************************/
1457
1458             if (gmx_mm_any_lt(rsq23,rcutoff2))
1459             {
1460
1461             /* REACTION-FIELD ELECTROSTATICS */
1462             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1463
1464             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1465
1466             fscal            = felec;
1467
1468             fscal            = _mm_and_pd(fscal,cutoff_mask);
1469
1470             /* Calculate temporary vectorial force */
1471             tx               = _mm_mul_pd(fscal,dx23);
1472             ty               = _mm_mul_pd(fscal,dy23);
1473             tz               = _mm_mul_pd(fscal,dz23);
1474
1475             /* Update vectorial force */
1476             fix2             = _mm_add_pd(fix2,tx);
1477             fiy2             = _mm_add_pd(fiy2,ty);
1478             fiz2             = _mm_add_pd(fiz2,tz);
1479
1480             fjx3             = _mm_add_pd(fjx3,tx);
1481             fjy3             = _mm_add_pd(fjy3,ty);
1482             fjz3             = _mm_add_pd(fjz3,tz);
1483
1484             }
1485
1486             /**************************
1487              * CALCULATE INTERACTIONS *
1488              **************************/
1489
1490             if (gmx_mm_any_lt(rsq31,rcutoff2))
1491             {
1492
1493             /* REACTION-FIELD ELECTROSTATICS */
1494             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1495
1496             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1497
1498             fscal            = felec;
1499
1500             fscal            = _mm_and_pd(fscal,cutoff_mask);
1501
1502             /* Calculate temporary vectorial force */
1503             tx               = _mm_mul_pd(fscal,dx31);
1504             ty               = _mm_mul_pd(fscal,dy31);
1505             tz               = _mm_mul_pd(fscal,dz31);
1506
1507             /* Update vectorial force */
1508             fix3             = _mm_add_pd(fix3,tx);
1509             fiy3             = _mm_add_pd(fiy3,ty);
1510             fiz3             = _mm_add_pd(fiz3,tz);
1511
1512             fjx1             = _mm_add_pd(fjx1,tx);
1513             fjy1             = _mm_add_pd(fjy1,ty);
1514             fjz1             = _mm_add_pd(fjz1,tz);
1515
1516             }
1517
1518             /**************************
1519              * CALCULATE INTERACTIONS *
1520              **************************/
1521
1522             if (gmx_mm_any_lt(rsq32,rcutoff2))
1523             {
1524
1525             /* REACTION-FIELD ELECTROSTATICS */
1526             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1527
1528             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1529
1530             fscal            = felec;
1531
1532             fscal            = _mm_and_pd(fscal,cutoff_mask);
1533
1534             /* Calculate temporary vectorial force */
1535             tx               = _mm_mul_pd(fscal,dx32);
1536             ty               = _mm_mul_pd(fscal,dy32);
1537             tz               = _mm_mul_pd(fscal,dz32);
1538
1539             /* Update vectorial force */
1540             fix3             = _mm_add_pd(fix3,tx);
1541             fiy3             = _mm_add_pd(fiy3,ty);
1542             fiz3             = _mm_add_pd(fiz3,tz);
1543
1544             fjx2             = _mm_add_pd(fjx2,tx);
1545             fjy2             = _mm_add_pd(fjy2,ty);
1546             fjz2             = _mm_add_pd(fjz2,tz);
1547
1548             }
1549
1550             /**************************
1551              * CALCULATE INTERACTIONS *
1552              **************************/
1553
1554             if (gmx_mm_any_lt(rsq33,rcutoff2))
1555             {
1556
1557             /* REACTION-FIELD ELECTROSTATICS */
1558             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1559
1560             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1561
1562             fscal            = felec;
1563
1564             fscal            = _mm_and_pd(fscal,cutoff_mask);
1565
1566             /* Calculate temporary vectorial force */
1567             tx               = _mm_mul_pd(fscal,dx33);
1568             ty               = _mm_mul_pd(fscal,dy33);
1569             tz               = _mm_mul_pd(fscal,dz33);
1570
1571             /* Update vectorial force */
1572             fix3             = _mm_add_pd(fix3,tx);
1573             fiy3             = _mm_add_pd(fiy3,ty);
1574             fiz3             = _mm_add_pd(fiz3,tz);
1575
1576             fjx3             = _mm_add_pd(fjx3,tx);
1577             fjy3             = _mm_add_pd(fjy3,ty);
1578             fjz3             = _mm_add_pd(fjz3,tz);
1579
1580             }
1581
1582             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1583
1584             /* Inner loop uses 270 flops */
1585         }
1586
1587         if(jidx<j_index_end)
1588         {
1589
1590             jnrA             = jjnr[jidx];
1591             j_coord_offsetA  = DIM*jnrA;
1592
1593             /* load j atom coordinates */
1594             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
1595                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1596
1597             /* Calculate displacement vector */
1598             dx11             = _mm_sub_pd(ix1,jx1);
1599             dy11             = _mm_sub_pd(iy1,jy1);
1600             dz11             = _mm_sub_pd(iz1,jz1);
1601             dx12             = _mm_sub_pd(ix1,jx2);
1602             dy12             = _mm_sub_pd(iy1,jy2);
1603             dz12             = _mm_sub_pd(iz1,jz2);
1604             dx13             = _mm_sub_pd(ix1,jx3);
1605             dy13             = _mm_sub_pd(iy1,jy3);
1606             dz13             = _mm_sub_pd(iz1,jz3);
1607             dx21             = _mm_sub_pd(ix2,jx1);
1608             dy21             = _mm_sub_pd(iy2,jy1);
1609             dz21             = _mm_sub_pd(iz2,jz1);
1610             dx22             = _mm_sub_pd(ix2,jx2);
1611             dy22             = _mm_sub_pd(iy2,jy2);
1612             dz22             = _mm_sub_pd(iz2,jz2);
1613             dx23             = _mm_sub_pd(ix2,jx3);
1614             dy23             = _mm_sub_pd(iy2,jy3);
1615             dz23             = _mm_sub_pd(iz2,jz3);
1616             dx31             = _mm_sub_pd(ix3,jx1);
1617             dy31             = _mm_sub_pd(iy3,jy1);
1618             dz31             = _mm_sub_pd(iz3,jz1);
1619             dx32             = _mm_sub_pd(ix3,jx2);
1620             dy32             = _mm_sub_pd(iy3,jy2);
1621             dz32             = _mm_sub_pd(iz3,jz2);
1622             dx33             = _mm_sub_pd(ix3,jx3);
1623             dy33             = _mm_sub_pd(iy3,jy3);
1624             dz33             = _mm_sub_pd(iz3,jz3);
1625
1626             /* Calculate squared distance and things based on it */
1627             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1628             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1629             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1630             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1631             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1632             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1633             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1634             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1635             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1636
1637             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1638             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1639             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1640             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1641             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1642             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1643             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1644             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1645             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1646
1647             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1648             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1649             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1650             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1651             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1652             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1653             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1654             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1655             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1656
1657             fjx1             = _mm_setzero_pd();
1658             fjy1             = _mm_setzero_pd();
1659             fjz1             = _mm_setzero_pd();
1660             fjx2             = _mm_setzero_pd();
1661             fjy2             = _mm_setzero_pd();
1662             fjz2             = _mm_setzero_pd();
1663             fjx3             = _mm_setzero_pd();
1664             fjy3             = _mm_setzero_pd();
1665             fjz3             = _mm_setzero_pd();
1666
1667             /**************************
1668              * CALCULATE INTERACTIONS *
1669              **************************/
1670
1671             if (gmx_mm_any_lt(rsq11,rcutoff2))
1672             {
1673
1674             /* REACTION-FIELD ELECTROSTATICS */
1675             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1676
1677             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1678
1679             fscal            = felec;
1680
1681             fscal            = _mm_and_pd(fscal,cutoff_mask);
1682
1683             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1684
1685             /* Calculate temporary vectorial force */
1686             tx               = _mm_mul_pd(fscal,dx11);
1687             ty               = _mm_mul_pd(fscal,dy11);
1688             tz               = _mm_mul_pd(fscal,dz11);
1689
1690             /* Update vectorial force */
1691             fix1             = _mm_add_pd(fix1,tx);
1692             fiy1             = _mm_add_pd(fiy1,ty);
1693             fiz1             = _mm_add_pd(fiz1,tz);
1694
1695             fjx1             = _mm_add_pd(fjx1,tx);
1696             fjy1             = _mm_add_pd(fjy1,ty);
1697             fjz1             = _mm_add_pd(fjz1,tz);
1698
1699             }
1700
1701             /**************************
1702              * CALCULATE INTERACTIONS *
1703              **************************/
1704
1705             if (gmx_mm_any_lt(rsq12,rcutoff2))
1706             {
1707
1708             /* REACTION-FIELD ELECTROSTATICS */
1709             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1710
1711             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1712
1713             fscal            = felec;
1714
1715             fscal            = _mm_and_pd(fscal,cutoff_mask);
1716
1717             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1718
1719             /* Calculate temporary vectorial force */
1720             tx               = _mm_mul_pd(fscal,dx12);
1721             ty               = _mm_mul_pd(fscal,dy12);
1722             tz               = _mm_mul_pd(fscal,dz12);
1723
1724             /* Update vectorial force */
1725             fix1             = _mm_add_pd(fix1,tx);
1726             fiy1             = _mm_add_pd(fiy1,ty);
1727             fiz1             = _mm_add_pd(fiz1,tz);
1728
1729             fjx2             = _mm_add_pd(fjx2,tx);
1730             fjy2             = _mm_add_pd(fjy2,ty);
1731             fjz2             = _mm_add_pd(fjz2,tz);
1732
1733             }
1734
1735             /**************************
1736              * CALCULATE INTERACTIONS *
1737              **************************/
1738
1739             if (gmx_mm_any_lt(rsq13,rcutoff2))
1740             {
1741
1742             /* REACTION-FIELD ELECTROSTATICS */
1743             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
1744
1745             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1746
1747             fscal            = felec;
1748
1749             fscal            = _mm_and_pd(fscal,cutoff_mask);
1750
1751             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1752
1753             /* Calculate temporary vectorial force */
1754             tx               = _mm_mul_pd(fscal,dx13);
1755             ty               = _mm_mul_pd(fscal,dy13);
1756             tz               = _mm_mul_pd(fscal,dz13);
1757
1758             /* Update vectorial force */
1759             fix1             = _mm_add_pd(fix1,tx);
1760             fiy1             = _mm_add_pd(fiy1,ty);
1761             fiz1             = _mm_add_pd(fiz1,tz);
1762
1763             fjx3             = _mm_add_pd(fjx3,tx);
1764             fjy3             = _mm_add_pd(fjy3,ty);
1765             fjz3             = _mm_add_pd(fjz3,tz);
1766
1767             }
1768
1769             /**************************
1770              * CALCULATE INTERACTIONS *
1771              **************************/
1772
1773             if (gmx_mm_any_lt(rsq21,rcutoff2))
1774             {
1775
1776             /* REACTION-FIELD ELECTROSTATICS */
1777             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1778
1779             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1780
1781             fscal            = felec;
1782
1783             fscal            = _mm_and_pd(fscal,cutoff_mask);
1784
1785             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1786
1787             /* Calculate temporary vectorial force */
1788             tx               = _mm_mul_pd(fscal,dx21);
1789             ty               = _mm_mul_pd(fscal,dy21);
1790             tz               = _mm_mul_pd(fscal,dz21);
1791
1792             /* Update vectorial force */
1793             fix2             = _mm_add_pd(fix2,tx);
1794             fiy2             = _mm_add_pd(fiy2,ty);
1795             fiz2             = _mm_add_pd(fiz2,tz);
1796
1797             fjx1             = _mm_add_pd(fjx1,tx);
1798             fjy1             = _mm_add_pd(fjy1,ty);
1799             fjz1             = _mm_add_pd(fjz1,tz);
1800
1801             }
1802
1803             /**************************
1804              * CALCULATE INTERACTIONS *
1805              **************************/
1806
1807             if (gmx_mm_any_lt(rsq22,rcutoff2))
1808             {
1809
1810             /* REACTION-FIELD ELECTROSTATICS */
1811             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1812
1813             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1814
1815             fscal            = felec;
1816
1817             fscal            = _mm_and_pd(fscal,cutoff_mask);
1818
1819             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1820
1821             /* Calculate temporary vectorial force */
1822             tx               = _mm_mul_pd(fscal,dx22);
1823             ty               = _mm_mul_pd(fscal,dy22);
1824             tz               = _mm_mul_pd(fscal,dz22);
1825
1826             /* Update vectorial force */
1827             fix2             = _mm_add_pd(fix2,tx);
1828             fiy2             = _mm_add_pd(fiy2,ty);
1829             fiz2             = _mm_add_pd(fiz2,tz);
1830
1831             fjx2             = _mm_add_pd(fjx2,tx);
1832             fjy2             = _mm_add_pd(fjy2,ty);
1833             fjz2             = _mm_add_pd(fjz2,tz);
1834
1835             }
1836
1837             /**************************
1838              * CALCULATE INTERACTIONS *
1839              **************************/
1840
1841             if (gmx_mm_any_lt(rsq23,rcutoff2))
1842             {
1843
1844             /* REACTION-FIELD ELECTROSTATICS */
1845             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1846
1847             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1848
1849             fscal            = felec;
1850
1851             fscal            = _mm_and_pd(fscal,cutoff_mask);
1852
1853             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1854
1855             /* Calculate temporary vectorial force */
1856             tx               = _mm_mul_pd(fscal,dx23);
1857             ty               = _mm_mul_pd(fscal,dy23);
1858             tz               = _mm_mul_pd(fscal,dz23);
1859
1860             /* Update vectorial force */
1861             fix2             = _mm_add_pd(fix2,tx);
1862             fiy2             = _mm_add_pd(fiy2,ty);
1863             fiz2             = _mm_add_pd(fiz2,tz);
1864
1865             fjx3             = _mm_add_pd(fjx3,tx);
1866             fjy3             = _mm_add_pd(fjy3,ty);
1867             fjz3             = _mm_add_pd(fjz3,tz);
1868
1869             }
1870
1871             /**************************
1872              * CALCULATE INTERACTIONS *
1873              **************************/
1874
1875             if (gmx_mm_any_lt(rsq31,rcutoff2))
1876             {
1877
1878             /* REACTION-FIELD ELECTROSTATICS */
1879             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1880
1881             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1882
1883             fscal            = felec;
1884
1885             fscal            = _mm_and_pd(fscal,cutoff_mask);
1886
1887             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1888
1889             /* Calculate temporary vectorial force */
1890             tx               = _mm_mul_pd(fscal,dx31);
1891             ty               = _mm_mul_pd(fscal,dy31);
1892             tz               = _mm_mul_pd(fscal,dz31);
1893
1894             /* Update vectorial force */
1895             fix3             = _mm_add_pd(fix3,tx);
1896             fiy3             = _mm_add_pd(fiy3,ty);
1897             fiz3             = _mm_add_pd(fiz3,tz);
1898
1899             fjx1             = _mm_add_pd(fjx1,tx);
1900             fjy1             = _mm_add_pd(fjy1,ty);
1901             fjz1             = _mm_add_pd(fjz1,tz);
1902
1903             }
1904
1905             /**************************
1906              * CALCULATE INTERACTIONS *
1907              **************************/
1908
1909             if (gmx_mm_any_lt(rsq32,rcutoff2))
1910             {
1911
1912             /* REACTION-FIELD ELECTROSTATICS */
1913             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1914
1915             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1916
1917             fscal            = felec;
1918
1919             fscal            = _mm_and_pd(fscal,cutoff_mask);
1920
1921             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1922
1923             /* Calculate temporary vectorial force */
1924             tx               = _mm_mul_pd(fscal,dx32);
1925             ty               = _mm_mul_pd(fscal,dy32);
1926             tz               = _mm_mul_pd(fscal,dz32);
1927
1928             /* Update vectorial force */
1929             fix3             = _mm_add_pd(fix3,tx);
1930             fiy3             = _mm_add_pd(fiy3,ty);
1931             fiz3             = _mm_add_pd(fiz3,tz);
1932
1933             fjx2             = _mm_add_pd(fjx2,tx);
1934             fjy2             = _mm_add_pd(fjy2,ty);
1935             fjz2             = _mm_add_pd(fjz2,tz);
1936
1937             }
1938
1939             /**************************
1940              * CALCULATE INTERACTIONS *
1941              **************************/
1942
1943             if (gmx_mm_any_lt(rsq33,rcutoff2))
1944             {
1945
1946             /* REACTION-FIELD ELECTROSTATICS */
1947             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1948
1949             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1950
1951             fscal            = felec;
1952
1953             fscal            = _mm_and_pd(fscal,cutoff_mask);
1954
1955             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1956
1957             /* Calculate temporary vectorial force */
1958             tx               = _mm_mul_pd(fscal,dx33);
1959             ty               = _mm_mul_pd(fscal,dy33);
1960             tz               = _mm_mul_pd(fscal,dz33);
1961
1962             /* Update vectorial force */
1963             fix3             = _mm_add_pd(fix3,tx);
1964             fiy3             = _mm_add_pd(fiy3,ty);
1965             fiz3             = _mm_add_pd(fiz3,tz);
1966
1967             fjx3             = _mm_add_pd(fjx3,tx);
1968             fjy3             = _mm_add_pd(fjy3,ty);
1969             fjz3             = _mm_add_pd(fjz3,tz);
1970
1971             }
1972
1973             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1974
1975             /* Inner loop uses 270 flops */
1976         }
1977
1978         /* End of innermost loop */
1979
1980         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1981                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1982
1983         /* Increment number of inner iterations */
1984         inneriter                  += j_index_end - j_index_start;
1985
1986         /* Outer loop uses 18 flops */
1987     }
1988
1989     /* Increment number of outer iterations */
1990     outeriter        += nri;
1991
1992     /* Update outer/inner flops */
1993
1994     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*270);
1995 }