src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_double/nb_kernel_ElecRFCut_VdwNone_GeomW4W4_sse4_1_double.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse4_1_double kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_sse4_1_double.h"
  34 #include "kernelutil_x86_sse4_1_double.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwNone_GeomW4W4_VF_sse4_1_double
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            None
  40  * Geometry:                   Water4-Water4
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRFCut_VdwNone_GeomW4W4_VF_sse4_1_double
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB;
  61     int              j_coord_offsetA,j_coord_offsetB;
  62     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  63     real             rcutoff_scalar;
  64     real             *shiftvec,*fshift,*x,*f;
  65     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  66     int              vdwioffset1;
  67     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  68     int              vdwioffset2;
  69     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  70     int              vdwioffset3;
  71     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  72     int              vdwjidx1A,vdwjidx1B;
  73     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  74     int              vdwjidx2A,vdwjidx2B;
  75     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  76     int              vdwjidx3A,vdwjidx3B;
  77     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  78     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  79     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  80     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  81     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  82     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  83     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  84     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  85     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
  86     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
  87     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  88     real             *charge;
  89     __m128d          dummy_mask,cutoff_mask;
  90     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
  91     __m128d          one     = _mm_set1_pd(1.0);
  92     __m128d          two     = _mm_set1_pd(2.0);
  93     x                = xx[0];
  94     f                = ff[0];
  95
  96     nri              = nlist->nri;
  97     iinr             = nlist->iinr;
  98     jindex           = nlist->jindex;
  99     jjnr             = nlist->jjnr;
 100     shiftidx         = nlist->shift;
 101     gid              = nlist->gid;
 102     shiftvec         = fr->shift_vec[0];
 103     fshift           = fr->fshift[0];
 104     facel            = _mm_set1_pd(fr->epsfac);
 105     charge           = mdatoms->chargeA;
 106     krf              = _mm_set1_pd(fr->ic->k_rf);
 107     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 108     crf              = _mm_set1_pd(fr->ic->c_rf);
 109
 110     /* Setup water-specific parameters */
 111     inr              = nlist->iinr[0];
 112     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 113     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 114     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 115
 116     jq1              = _mm_set1_pd(charge[inr+1]);
 117     jq2              = _mm_set1_pd(charge[inr+2]);
 118     jq3              = _mm_set1_pd(charge[inr+3]);
 119     qq11             = _mm_mul_pd(iq1,jq1);
 120     qq12             = _mm_mul_pd(iq1,jq2);
 121     qq13             = _mm_mul_pd(iq1,jq3);
 122     qq21             = _mm_mul_pd(iq2,jq1);
 123     qq22             = _mm_mul_pd(iq2,jq2);
 124     qq23             = _mm_mul_pd(iq2,jq3);
 125     qq31             = _mm_mul_pd(iq3,jq1);
 126     qq32             = _mm_mul_pd(iq3,jq2);
 127     qq33             = _mm_mul_pd(iq3,jq3);
 128
 129     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 130     rcutoff_scalar   = fr->rcoulomb;
 131     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 132     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 133
 134     /* Avoid stupid compiler warnings */
 135     jnrA = jnrB = 0;
 136     j_coord_offsetA = 0;
 137     j_coord_offsetB = 0;
 138
 139     outeriter        = 0;
 140     inneriter        = 0;
 141
 142     /* Start outer loop over neighborlists */
 143     for(iidx=0; iidx<nri; iidx++)
 144     {
 145         /* Load shift vector for this list */
 146         i_shift_offset   = DIM*shiftidx[iidx];
 147
 148         /* Load limits for loop over neighbors */
 149         j_index_start    = jindex[iidx];
 150         j_index_end      = jindex[iidx+1];
 151
 152         /* Get outer coordinate index */
 153         inr              = iinr[iidx];
 154         i_coord_offset   = DIM*inr;
 155
 156         /* Load i particle coords and add shift vector */
 157         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 158                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 159
 160         fix1             = _mm_setzero_pd();
 161         fiy1             = _mm_setzero_pd();
 162         fiz1             = _mm_setzero_pd();
 163         fix2             = _mm_setzero_pd();
 164         fiy2             = _mm_setzero_pd();
 165         fiz2             = _mm_setzero_pd();
 166         fix3             = _mm_setzero_pd();
 167         fiy3             = _mm_setzero_pd();
 168         fiz3             = _mm_setzero_pd();
 169
 170         /* Reset potential sums */
 171         velecsum         = _mm_setzero_pd();
 172
 173         /* Start inner kernel loop */
 174         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 175         {
 176
 177             /* Get j neighbor index, and coordinate index */
 178             jnrA             = jjnr[jidx];
 179             jnrB             = jjnr[jidx+1];
 180             j_coord_offsetA  = DIM*jnrA;
 181             j_coord_offsetB  = DIM*jnrB;
 182
 183             /* load j atom coordinates */
 184             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 185                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 186
 187             /* Calculate displacement vector */
 188             dx11             = _mm_sub_pd(ix1,jx1);
 189             dy11             = _mm_sub_pd(iy1,jy1);
 190             dz11             = _mm_sub_pd(iz1,jz1);
 191             dx12             = _mm_sub_pd(ix1,jx2);
 192             dy12             = _mm_sub_pd(iy1,jy2);
 193             dz12             = _mm_sub_pd(iz1,jz2);
 194             dx13             = _mm_sub_pd(ix1,jx3);
 195             dy13             = _mm_sub_pd(iy1,jy3);
 196             dz13             = _mm_sub_pd(iz1,jz3);
 197             dx21             = _mm_sub_pd(ix2,jx1);
 198             dy21             = _mm_sub_pd(iy2,jy1);
 199             dz21             = _mm_sub_pd(iz2,jz1);
 200             dx22             = _mm_sub_pd(ix2,jx2);
 201             dy22             = _mm_sub_pd(iy2,jy2);
 202             dz22             = _mm_sub_pd(iz2,jz2);
 203             dx23             = _mm_sub_pd(ix2,jx3);
 204             dy23             = _mm_sub_pd(iy2,jy3);
 205             dz23             = _mm_sub_pd(iz2,jz3);
 206             dx31             = _mm_sub_pd(ix3,jx1);
 207             dy31             = _mm_sub_pd(iy3,jy1);
 208             dz31             = _mm_sub_pd(iz3,jz1);
 209             dx32             = _mm_sub_pd(ix3,jx2);
 210             dy32             = _mm_sub_pd(iy3,jy2);
 211             dz32             = _mm_sub_pd(iz3,jz2);
 212             dx33             = _mm_sub_pd(ix3,jx3);
 213             dy33             = _mm_sub_pd(iy3,jy3);
 214             dz33             = _mm_sub_pd(iz3,jz3);
 215
 216             /* Calculate squared distance and things based on it */
 217             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 218             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 219             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 220             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 221             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 222             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 223             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 224             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 225             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 226
 227             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 228             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 229             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 230             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 231             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 232             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 233             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 234             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 235             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 236
 237             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 238             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 239             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 240             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 241             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 242             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 243             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 244             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 245             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 246
 247             fjx1             = _mm_setzero_pd();
 248             fjy1             = _mm_setzero_pd();
 249             fjz1             = _mm_setzero_pd();
 250             fjx2             = _mm_setzero_pd();
 251             fjy2             = _mm_setzero_pd();
 252             fjz2             = _mm_setzero_pd();
 253             fjx3             = _mm_setzero_pd();
 254             fjy3             = _mm_setzero_pd();
 255             fjz3             = _mm_setzero_pd();
 256
 257             /**************************
 258              * CALCULATE INTERACTIONS *
 259              **************************/
 260
 261             if (gmx_mm_any_lt(rsq11,rcutoff2))
 262             {
 263
 264             /* REACTION-FIELD ELECTROSTATICS */
 265             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 266             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 267
 268             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 269
 270             /* Update potential sum for this i atom from the interaction with this j atom. */
 271             velec            = _mm_and_pd(velec,cutoff_mask);
 272             velecsum         = _mm_add_pd(velecsum,velec);
 273
 274             fscal            = felec;
 275
 276             fscal            = _mm_and_pd(fscal,cutoff_mask);
 277
 278             /* Calculate temporary vectorial force */
 279             tx               = _mm_mul_pd(fscal,dx11);
 280             ty               = _mm_mul_pd(fscal,dy11);
 281             tz               = _mm_mul_pd(fscal,dz11);
 282
 283             /* Update vectorial force */
 284             fix1             = _mm_add_pd(fix1,tx);
 285             fiy1             = _mm_add_pd(fiy1,ty);
 286             fiz1             = _mm_add_pd(fiz1,tz);
 287
 288             fjx1             = _mm_add_pd(fjx1,tx);
 289             fjy1             = _mm_add_pd(fjy1,ty);
 290             fjz1             = _mm_add_pd(fjz1,tz);
 291
 292             }
 293
 294             /**************************
 295              * CALCULATE INTERACTIONS *
 296              **************************/
 297
 298             if (gmx_mm_any_lt(rsq12,rcutoff2))
 299             {
 300
 301             /* REACTION-FIELD ELECTROSTATICS */
 302             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 303             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 304
 305             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 306
 307             /* Update potential sum for this i atom from the interaction with this j atom. */
 308             velec            = _mm_and_pd(velec,cutoff_mask);
 309             velecsum         = _mm_add_pd(velecsum,velec);
 310
 311             fscal            = felec;
 312
 313             fscal            = _mm_and_pd(fscal,cutoff_mask);
 314
 315             /* Calculate temporary vectorial force */
 316             tx               = _mm_mul_pd(fscal,dx12);
 317             ty               = _mm_mul_pd(fscal,dy12);
 318             tz               = _mm_mul_pd(fscal,dz12);
 319
 320             /* Update vectorial force */
 321             fix1             = _mm_add_pd(fix1,tx);
 322             fiy1             = _mm_add_pd(fiy1,ty);
 323             fiz1             = _mm_add_pd(fiz1,tz);
 324
 325             fjx2             = _mm_add_pd(fjx2,tx);
 326             fjy2             = _mm_add_pd(fjy2,ty);
 327             fjz2             = _mm_add_pd(fjz2,tz);
 328
 329             }
 330
 331             /**************************
 332              * CALCULATE INTERACTIONS *
 333              **************************/
 334
 335             if (gmx_mm_any_lt(rsq13,rcutoff2))
 336             {
 337
 338             /* REACTION-FIELD ELECTROSTATICS */
 339             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
 340             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
 341
 342             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 343
 344             /* Update potential sum for this i atom from the interaction with this j atom. */
 345             velec            = _mm_and_pd(velec,cutoff_mask);
 346             velecsum         = _mm_add_pd(velecsum,velec);
 347
 348             fscal            = felec;
 349
 350             fscal            = _mm_and_pd(fscal,cutoff_mask);
 351
 352             /* Calculate temporary vectorial force */
 353             tx               = _mm_mul_pd(fscal,dx13);
 354             ty               = _mm_mul_pd(fscal,dy13);
 355             tz               = _mm_mul_pd(fscal,dz13);
 356
 357             /* Update vectorial force */
 358             fix1             = _mm_add_pd(fix1,tx);
 359             fiy1             = _mm_add_pd(fiy1,ty);
 360             fiz1             = _mm_add_pd(fiz1,tz);
 361
 362             fjx3             = _mm_add_pd(fjx3,tx);
 363             fjy3             = _mm_add_pd(fjy3,ty);
 364             fjz3             = _mm_add_pd(fjz3,tz);
 365
 366             }
 367
 368             /**************************
 369              * CALCULATE INTERACTIONS *
 370              **************************/
 371
 372             if (gmx_mm_any_lt(rsq21,rcutoff2))
 373             {
 374
 375             /* REACTION-FIELD ELECTROSTATICS */
 376             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 377             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 378
 379             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 380
 381             /* Update potential sum for this i atom from the interaction with this j atom. */
 382             velec            = _mm_and_pd(velec,cutoff_mask);
 383             velecsum         = _mm_add_pd(velecsum,velec);
 384
 385             fscal            = felec;
 386
 387             fscal            = _mm_and_pd(fscal,cutoff_mask);
 388
 389             /* Calculate temporary vectorial force */
 390             tx               = _mm_mul_pd(fscal,dx21);
 391             ty               = _mm_mul_pd(fscal,dy21);
 392             tz               = _mm_mul_pd(fscal,dz21);
 393
 394             /* Update vectorial force */
 395             fix2             = _mm_add_pd(fix2,tx);
 396             fiy2             = _mm_add_pd(fiy2,ty);
 397             fiz2             = _mm_add_pd(fiz2,tz);
 398
 399             fjx1             = _mm_add_pd(fjx1,tx);
 400             fjy1             = _mm_add_pd(fjy1,ty);
 401             fjz1             = _mm_add_pd(fjz1,tz);
 402
 403             }
 404
 405             /**************************
 406              * CALCULATE INTERACTIONS *
 407              **************************/
 408
 409             if (gmx_mm_any_lt(rsq22,rcutoff2))
 410             {
 411
 412             /* REACTION-FIELD ELECTROSTATICS */
 413             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 414             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 415
 416             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 417
 418             /* Update potential sum for this i atom from the interaction with this j atom. */
 419             velec            = _mm_and_pd(velec,cutoff_mask);
 420             velecsum         = _mm_add_pd(velecsum,velec);
 421
 422             fscal            = felec;
 423
 424             fscal            = _mm_and_pd(fscal,cutoff_mask);
 425
 426             /* Calculate temporary vectorial force */
 427             tx               = _mm_mul_pd(fscal,dx22);
 428             ty               = _mm_mul_pd(fscal,dy22);
 429             tz               = _mm_mul_pd(fscal,dz22);
 430
 431             /* Update vectorial force */
 432             fix2             = _mm_add_pd(fix2,tx);
 433             fiy2             = _mm_add_pd(fiy2,ty);
 434             fiz2             = _mm_add_pd(fiz2,tz);
 435
 436             fjx2             = _mm_add_pd(fjx2,tx);
 437             fjy2             = _mm_add_pd(fjy2,ty);
 438             fjz2             = _mm_add_pd(fjz2,tz);
 439
 440             }
 441
 442             /**************************
 443              * CALCULATE INTERACTIONS *
 444              **************************/
 445
 446             if (gmx_mm_any_lt(rsq23,rcutoff2))
 447             {
 448
 449             /* REACTION-FIELD ELECTROSTATICS */
 450             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
 451             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
 452
 453             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 454
 455             /* Update potential sum for this i atom from the interaction with this j atom. */
 456             velec            = _mm_and_pd(velec,cutoff_mask);
 457             velecsum         = _mm_add_pd(velecsum,velec);
 458
 459             fscal            = felec;
 460
 461             fscal            = _mm_and_pd(fscal,cutoff_mask);
 462
 463             /* Calculate temporary vectorial force */
 464             tx               = _mm_mul_pd(fscal,dx23);
 465             ty               = _mm_mul_pd(fscal,dy23);
 466             tz               = _mm_mul_pd(fscal,dz23);
 467
 468             /* Update vectorial force */
 469             fix2             = _mm_add_pd(fix2,tx);
 470             fiy2             = _mm_add_pd(fiy2,ty);
 471             fiz2             = _mm_add_pd(fiz2,tz);
 472
 473             fjx3             = _mm_add_pd(fjx3,tx);
 474             fjy3             = _mm_add_pd(fjy3,ty);
 475             fjz3             = _mm_add_pd(fjz3,tz);
 476
 477             }
 478
 479             /**************************
 480              * CALCULATE INTERACTIONS *
 481              **************************/
 482
 483             if (gmx_mm_any_lt(rsq31,rcutoff2))
 484             {
 485
 486             /* REACTION-FIELD ELECTROSTATICS */
 487             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
 488             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
 489
 490             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
 491
 492             /* Update potential sum for this i atom from the interaction with this j atom. */
 493             velec            = _mm_and_pd(velec,cutoff_mask);
 494             velecsum         = _mm_add_pd(velecsum,velec);
 495
 496             fscal            = felec;
 497
 498             fscal            = _mm_and_pd(fscal,cutoff_mask);
 499
 500             /* Calculate temporary vectorial force */
 501             tx               = _mm_mul_pd(fscal,dx31);
 502             ty               = _mm_mul_pd(fscal,dy31);
 503             tz               = _mm_mul_pd(fscal,dz31);
 504
 505             /* Update vectorial force */
 506             fix3             = _mm_add_pd(fix3,tx);
 507             fiy3             = _mm_add_pd(fiy3,ty);
 508             fiz3             = _mm_add_pd(fiz3,tz);
 509
 510             fjx1             = _mm_add_pd(fjx1,tx);
 511             fjy1             = _mm_add_pd(fjy1,ty);
 512             fjz1             = _mm_add_pd(fjz1,tz);
 513
 514             }
 515
 516             /**************************
 517              * CALCULATE INTERACTIONS *
 518              **************************/
 519
 520             if (gmx_mm_any_lt(rsq32,rcutoff2))
 521             {
 522
 523             /* REACTION-FIELD ELECTROSTATICS */
 524             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
 525             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
 526
 527             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
 528
 529             /* Update potential sum for this i atom from the interaction with this j atom. */
 530             velec            = _mm_and_pd(velec,cutoff_mask);
 531             velecsum         = _mm_add_pd(velecsum,velec);
 532
 533             fscal            = felec;
 534
 535             fscal            = _mm_and_pd(fscal,cutoff_mask);
 536
 537             /* Calculate temporary vectorial force */
 538             tx               = _mm_mul_pd(fscal,dx32);
 539             ty               = _mm_mul_pd(fscal,dy32);
 540             tz               = _mm_mul_pd(fscal,dz32);
 541
 542             /* Update vectorial force */
 543             fix3             = _mm_add_pd(fix3,tx);
 544             fiy3             = _mm_add_pd(fiy3,ty);
 545             fiz3             = _mm_add_pd(fiz3,tz);
 546
 547             fjx2             = _mm_add_pd(fjx2,tx);
 548             fjy2             = _mm_add_pd(fjy2,ty);
 549             fjz2             = _mm_add_pd(fjz2,tz);
 550
 551             }
 552
 553             /**************************
 554              * CALCULATE INTERACTIONS *
 555              **************************/
 556
 557             if (gmx_mm_any_lt(rsq33,rcutoff2))
 558             {
 559
 560             /* REACTION-FIELD ELECTROSTATICS */
 561             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
 562             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
 563
 564             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
 565
 566             /* Update potential sum for this i atom from the interaction with this j atom. */
 567             velec            = _mm_and_pd(velec,cutoff_mask);
 568             velecsum         = _mm_add_pd(velecsum,velec);
 569
 570             fscal            = felec;
 571
 572             fscal            = _mm_and_pd(fscal,cutoff_mask);
 573
 574             /* Calculate temporary vectorial force */
 575             tx               = _mm_mul_pd(fscal,dx33);
 576             ty               = _mm_mul_pd(fscal,dy33);
 577             tz               = _mm_mul_pd(fscal,dz33);
 578
 579             /* Update vectorial force */
 580             fix3             = _mm_add_pd(fix3,tx);
 581             fiy3             = _mm_add_pd(fiy3,ty);
 582             fiz3             = _mm_add_pd(fiz3,tz);
 583
 584             fjx3             = _mm_add_pd(fjx3,tx);
 585             fjy3             = _mm_add_pd(fjy3,ty);
 586             fjz3             = _mm_add_pd(fjz3,tz);
 587
 588             }
 589
 590             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);
 591
 592             /* Inner loop uses 324 flops */
 593         }
 594
 595         if(jidx<j_index_end)
 596         {
 597
 598             jnrA             = jjnr[jidx];
 599             j_coord_offsetA  = DIM*jnrA;
 600
 601             /* load j atom coordinates */
 602             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
 603                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 604
 605             /* Calculate displacement vector */
 606             dx11             = _mm_sub_pd(ix1,jx1);
 607             dy11             = _mm_sub_pd(iy1,jy1);
 608             dz11             = _mm_sub_pd(iz1,jz1);
 609             dx12             = _mm_sub_pd(ix1,jx2);
 610             dy12             = _mm_sub_pd(iy1,jy2);
 611             dz12             = _mm_sub_pd(iz1,jz2);
 612             dx13             = _mm_sub_pd(ix1,jx3);
 613             dy13             = _mm_sub_pd(iy1,jy3);
 614             dz13             = _mm_sub_pd(iz1,jz3);
 615             dx21             = _mm_sub_pd(ix2,jx1);
 616             dy21             = _mm_sub_pd(iy2,jy1);
 617             dz21             = _mm_sub_pd(iz2,jz1);
 618             dx22             = _mm_sub_pd(ix2,jx2);
 619             dy22             = _mm_sub_pd(iy2,jy2);
 620             dz22             = _mm_sub_pd(iz2,jz2);
 621             dx23             = _mm_sub_pd(ix2,jx3);
 622             dy23             = _mm_sub_pd(iy2,jy3);
 623             dz23             = _mm_sub_pd(iz2,jz3);
 624             dx31             = _mm_sub_pd(ix3,jx1);
 625             dy31             = _mm_sub_pd(iy3,jy1);
 626             dz31             = _mm_sub_pd(iz3,jz1);
 627             dx32             = _mm_sub_pd(ix3,jx2);
 628             dy32             = _mm_sub_pd(iy3,jy2);
 629             dz32             = _mm_sub_pd(iz3,jz2);
 630             dx33             = _mm_sub_pd(ix3,jx3);
 631             dy33             = _mm_sub_pd(iy3,jy3);
 632             dz33             = _mm_sub_pd(iz3,jz3);
 633
 634             /* Calculate squared distance and things based on it */
 635             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 636             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 637             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 638             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 639             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 640             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 641             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 642             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 643             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 644
 645             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 646             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 647             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 648             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 649             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 650             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 651             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 652             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 653             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 654
 655             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 656             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 657             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 658             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 659             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 660             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 661             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 662             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 663             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 664
 665             fjx1             = _mm_setzero_pd();
 666             fjy1             = _mm_setzero_pd();
 667             fjz1             = _mm_setzero_pd();
 668             fjx2             = _mm_setzero_pd();
 669             fjy2             = _mm_setzero_pd();
 670             fjz2             = _mm_setzero_pd();
 671             fjx3             = _mm_setzero_pd();
 672             fjy3             = _mm_setzero_pd();
 673             fjz3             = _mm_setzero_pd();
 674
 675             /**************************
 676              * CALCULATE INTERACTIONS *
 677              **************************/
 678
 679             if (gmx_mm_any_lt(rsq11,rcutoff2))
 680             {
 681
 682             /* REACTION-FIELD ELECTROSTATICS */
 683             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 684             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 685
 686             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 687
 688             /* Update potential sum for this i atom from the interaction with this j atom. */
 689             velec            = _mm_and_pd(velec,cutoff_mask);
 690             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 691             velecsum         = _mm_add_pd(velecsum,velec);
 692
 693             fscal            = felec;
 694
 695             fscal            = _mm_and_pd(fscal,cutoff_mask);
 696
 697             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 698
 699             /* Calculate temporary vectorial force */
 700             tx               = _mm_mul_pd(fscal,dx11);
 701             ty               = _mm_mul_pd(fscal,dy11);
 702             tz               = _mm_mul_pd(fscal,dz11);
 703
 704             /* Update vectorial force */
 705             fix1             = _mm_add_pd(fix1,tx);
 706             fiy1             = _mm_add_pd(fiy1,ty);
 707             fiz1             = _mm_add_pd(fiz1,tz);
 708
 709             fjx1             = _mm_add_pd(fjx1,tx);
 710             fjy1             = _mm_add_pd(fjy1,ty);
 711             fjz1             = _mm_add_pd(fjz1,tz);
 712
 713             }
 714
 715             /**************************
 716              * CALCULATE INTERACTIONS *
 717              **************************/
 718
 719             if (gmx_mm_any_lt(rsq12,rcutoff2))
 720             {
 721
 722             /* REACTION-FIELD ELECTROSTATICS */
 723             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 724             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 725
 726             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 727
 728             /* Update potential sum for this i atom from the interaction with this j atom. */
 729             velec            = _mm_and_pd(velec,cutoff_mask);
 730             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 731             velecsum         = _mm_add_pd(velecsum,velec);
 732
 733             fscal            = felec;
 734
 735             fscal            = _mm_and_pd(fscal,cutoff_mask);
 736
 737             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 738
 739             /* Calculate temporary vectorial force */
 740             tx               = _mm_mul_pd(fscal,dx12);
 741             ty               = _mm_mul_pd(fscal,dy12);
 742             tz               = _mm_mul_pd(fscal,dz12);
 743
 744             /* Update vectorial force */
 745             fix1             = _mm_add_pd(fix1,tx);
 746             fiy1             = _mm_add_pd(fiy1,ty);
 747             fiz1             = _mm_add_pd(fiz1,tz);
 748
 749             fjx2             = _mm_add_pd(fjx2,tx);
 750             fjy2             = _mm_add_pd(fjy2,ty);
 751             fjz2             = _mm_add_pd(fjz2,tz);
 752
 753             }
 754
 755             /**************************
 756              * CALCULATE INTERACTIONS *
 757              **************************/
 758
 759             if (gmx_mm_any_lt(rsq13,rcutoff2))
 760             {
 761
 762             /* REACTION-FIELD ELECTROSTATICS */
 763             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
 764             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
 765
 766             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 767
 768             /* Update potential sum for this i atom from the interaction with this j atom. */
 769             velec            = _mm_and_pd(velec,cutoff_mask);
 770             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 771             velecsum         = _mm_add_pd(velecsum,velec);
 772
 773             fscal            = felec;
 774
 775             fscal            = _mm_and_pd(fscal,cutoff_mask);
 776
 777             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 778
 779             /* Calculate temporary vectorial force */
 780             tx               = _mm_mul_pd(fscal,dx13);
 781             ty               = _mm_mul_pd(fscal,dy13);
 782             tz               = _mm_mul_pd(fscal,dz13);
 783
 784             /* Update vectorial force */
 785             fix1             = _mm_add_pd(fix1,tx);
 786             fiy1             = _mm_add_pd(fiy1,ty);
 787             fiz1             = _mm_add_pd(fiz1,tz);
 788
 789             fjx3             = _mm_add_pd(fjx3,tx);
 790             fjy3             = _mm_add_pd(fjy3,ty);
 791             fjz3             = _mm_add_pd(fjz3,tz);
 792
 793             }
 794
 795             /**************************
 796              * CALCULATE INTERACTIONS *
 797              **************************/
 798
 799             if (gmx_mm_any_lt(rsq21,rcutoff2))
 800             {
 801
 802             /* REACTION-FIELD ELECTROSTATICS */
 803             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 804             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 805
 806             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 807
 808             /* Update potential sum for this i atom from the interaction with this j atom. */
 809             velec            = _mm_and_pd(velec,cutoff_mask);
 810             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 811             velecsum         = _mm_add_pd(velecsum,velec);
 812
 813             fscal            = felec;
 814
 815             fscal            = _mm_and_pd(fscal,cutoff_mask);
 816
 817             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 818
 819             /* Calculate temporary vectorial force */
 820             tx               = _mm_mul_pd(fscal,dx21);
 821             ty               = _mm_mul_pd(fscal,dy21);
 822             tz               = _mm_mul_pd(fscal,dz21);
 823
 824             /* Update vectorial force */
 825             fix2             = _mm_add_pd(fix2,tx);
 826             fiy2             = _mm_add_pd(fiy2,ty);
 827             fiz2             = _mm_add_pd(fiz2,tz);
 828
 829             fjx1             = _mm_add_pd(fjx1,tx);
 830             fjy1             = _mm_add_pd(fjy1,ty);
 831             fjz1             = _mm_add_pd(fjz1,tz);
 832
 833             }
 834
 835             /**************************
 836              * CALCULATE INTERACTIONS *
 837              **************************/
 838
 839             if (gmx_mm_any_lt(rsq22,rcutoff2))
 840             {
 841
 842             /* REACTION-FIELD ELECTROSTATICS */
 843             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 844             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 845
 846             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 847
 848             /* Update potential sum for this i atom from the interaction with this j atom. */
 849             velec            = _mm_and_pd(velec,cutoff_mask);
 850             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 851             velecsum         = _mm_add_pd(velecsum,velec);
 852
 853             fscal            = felec;
 854
 855             fscal            = _mm_and_pd(fscal,cutoff_mask);
 856
 857             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 858
 859             /* Calculate temporary vectorial force */
 860             tx               = _mm_mul_pd(fscal,dx22);
 861             ty               = _mm_mul_pd(fscal,dy22);
 862             tz               = _mm_mul_pd(fscal,dz22);
 863
 864             /* Update vectorial force */
 865             fix2             = _mm_add_pd(fix2,tx);
 866             fiy2             = _mm_add_pd(fiy2,ty);
 867             fiz2             = _mm_add_pd(fiz2,tz);
 868
 869             fjx2             = _mm_add_pd(fjx2,tx);
 870             fjy2             = _mm_add_pd(fjy2,ty);
 871             fjz2             = _mm_add_pd(fjz2,tz);
 872
 873             }
 874
 875             /**************************
 876              * CALCULATE INTERACTIONS *
 877              **************************/
 878
 879             if (gmx_mm_any_lt(rsq23,rcutoff2))
 880             {
 881
 882             /* REACTION-FIELD ELECTROSTATICS */
 883             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
 884             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
 885
 886             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 887
 888             /* Update potential sum for this i atom from the interaction with this j atom. */
 889             velec            = _mm_and_pd(velec,cutoff_mask);
 890             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 891             velecsum         = _mm_add_pd(velecsum,velec);
 892
 893             fscal            = felec;
 894
 895             fscal            = _mm_and_pd(fscal,cutoff_mask);
 896
 897             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 898
 899             /* Calculate temporary vectorial force */
 900             tx               = _mm_mul_pd(fscal,dx23);
 901             ty               = _mm_mul_pd(fscal,dy23);
 902             tz               = _mm_mul_pd(fscal,dz23);
 903
 904             /* Update vectorial force */
 905             fix2             = _mm_add_pd(fix2,tx);
 906             fiy2             = _mm_add_pd(fiy2,ty);
 907             fiz2             = _mm_add_pd(fiz2,tz);
 908
 909             fjx3             = _mm_add_pd(fjx3,tx);
 910             fjy3             = _mm_add_pd(fjy3,ty);
 911             fjz3             = _mm_add_pd(fjz3,tz);
 912
 913             }
 914
 915             /**************************
 916              * CALCULATE INTERACTIONS *
 917              **************************/
 918
 919             if (gmx_mm_any_lt(rsq31,rcutoff2))
 920             {
 921
 922             /* REACTION-FIELD ELECTROSTATICS */
 923             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
 924             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
 925
 926             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
 927
 928             /* Update potential sum for this i atom from the interaction with this j atom. */
 929             velec            = _mm_and_pd(velec,cutoff_mask);
 930             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 931             velecsum         = _mm_add_pd(velecsum,velec);
 932
 933             fscal            = felec;
 934
 935             fscal            = _mm_and_pd(fscal,cutoff_mask);
 936
 937             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 938
 939             /* Calculate temporary vectorial force */
 940             tx               = _mm_mul_pd(fscal,dx31);
 941             ty               = _mm_mul_pd(fscal,dy31);
 942             tz               = _mm_mul_pd(fscal,dz31);
 943
 944             /* Update vectorial force */
 945             fix3             = _mm_add_pd(fix3,tx);
 946             fiy3             = _mm_add_pd(fiy3,ty);
 947             fiz3             = _mm_add_pd(fiz3,tz);
 948
 949             fjx1             = _mm_add_pd(fjx1,tx);
 950             fjy1             = _mm_add_pd(fjy1,ty);
 951             fjz1             = _mm_add_pd(fjz1,tz);
 952
 953             }
 954
 955             /**************************
 956              * CALCULATE INTERACTIONS *
 957              **************************/
 958
 959             if (gmx_mm_any_lt(rsq32,rcutoff2))
 960             {
 961
 962             /* REACTION-FIELD ELECTROSTATICS */
 963             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
 964             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
 965
 966             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
 967
 968             /* Update potential sum for this i atom from the interaction with this j atom. */
 969             velec            = _mm_and_pd(velec,cutoff_mask);
 970             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 971             velecsum         = _mm_add_pd(velecsum,velec);
 972
 973             fscal            = felec;
 974
 975             fscal            = _mm_and_pd(fscal,cutoff_mask);
 976
 977             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 978
 979             /* Calculate temporary vectorial force */
 980             tx               = _mm_mul_pd(fscal,dx32);
 981             ty               = _mm_mul_pd(fscal,dy32);
 982             tz               = _mm_mul_pd(fscal,dz32);
 983
 984             /* Update vectorial force */
 985             fix3             = _mm_add_pd(fix3,tx);
 986             fiy3             = _mm_add_pd(fiy3,ty);
 987             fiz3             = _mm_add_pd(fiz3,tz);
 988
 989             fjx2             = _mm_add_pd(fjx2,tx);
 990             fjy2             = _mm_add_pd(fjy2,ty);
 991             fjz2             = _mm_add_pd(fjz2,tz);
 992
 993             }
 994
 995             /**************************
 996              * CALCULATE INTERACTIONS *
 997              **************************/
 998
 999             if (gmx_mm_any_lt(rsq33,rcutoff2))
1000             {
1001
1002             /* REACTION-FIELD ELECTROSTATICS */
1003             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
1004             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1005
1006             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1007
1008             /* Update potential sum for this i atom from the interaction with this j atom. */
1009             velec            = _mm_and_pd(velec,cutoff_mask);
1010             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1011             velecsum         = _mm_add_pd(velecsum,velec);
1012
1013             fscal            = felec;
1014
1015             fscal            = _mm_and_pd(fscal,cutoff_mask);
1016
1017             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1018
1019             /* Calculate temporary vectorial force */
1020             tx               = _mm_mul_pd(fscal,dx33);
1021             ty               = _mm_mul_pd(fscal,dy33);
1022             tz               = _mm_mul_pd(fscal,dz33);
1023
1024             /* Update vectorial force */
1025             fix3             = _mm_add_pd(fix3,tx);
1026             fiy3             = _mm_add_pd(fiy3,ty);
1027             fiz3             = _mm_add_pd(fiz3,tz);
1028
1029             fjx3             = _mm_add_pd(fjx3,tx);
1030             fjy3             = _mm_add_pd(fjy3,ty);
1031             fjz3             = _mm_add_pd(fjz3,tz);
1032
1033             }
1034
1035             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1036
1037             /* Inner loop uses 324 flops */
1038         }
1039
1040         /* End of innermost loop */
1041
1042         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1043                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1044
1045         ggid                        = gid[iidx];
1046         /* Update potential energies */
1047         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1048
1049         /* Increment number of inner iterations */
1050         inneriter                  += j_index_end - j_index_start;
1051
1052         /* Outer loop uses 19 flops */
1053     }
1054
1055     /* Increment number of outer iterations */
1056     outeriter        += nri;
1057
1058     /* Update outer/inner flops */
1059
1060     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*324);
1061 }
1062 /*
1063  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwNone_GeomW4W4_F_sse4_1_double
1064  * Electrostatics interaction: ReactionField
1065  * VdW interaction:            None
1066  * Geometry:                   Water4-Water4
1067  * Calculate force/pot:        Force
1068  */
1069 void
1070 nb_kernel_ElecRFCut_VdwNone_GeomW4W4_F_sse4_1_double
1071                     (t_nblist * gmx_restrict                nlist,
1072                      rvec * gmx_restrict                    xx,
1073                      rvec * gmx_restrict                    ff,
1074                      t_forcerec * gmx_restrict              fr,
1075                      t_mdatoms * gmx_restrict               mdatoms,
1076                      nb_kernel_data_t * gmx_restrict        kernel_data,
1077                      t_nrnb * gmx_restrict                  nrnb)
1078 {
1079     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1080      * just 0 for non-waters.
1081      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1082      * jnr indices corresponding to data put in the four positions in the SIMD register.
1083      */
1084     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1085     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1086     int              jnrA,jnrB;
1087     int              j_coord_offsetA,j_coord_offsetB;
1088     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1089     real             rcutoff_scalar;
1090     real             *shiftvec,*fshift,*x,*f;
1091     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1092     int              vdwioffset1;
1093     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1094     int              vdwioffset2;
1095     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1096     int              vdwioffset3;
1097     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1098     int              vdwjidx1A,vdwjidx1B;
1099     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1100     int              vdwjidx2A,vdwjidx2B;
1101     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1102     int              vdwjidx3A,vdwjidx3B;
1103     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1104     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1105     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1106     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1107     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1108     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1109     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1110     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1111     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1112     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1113     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1114     real             *charge;
1115     __m128d          dummy_mask,cutoff_mask;
1116     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1117     __m128d          one     = _mm_set1_pd(1.0);
1118     __m128d          two     = _mm_set1_pd(2.0);
1119     x                = xx[0];
1120     f                = ff[0];
1121
1122     nri              = nlist->nri;
1123     iinr             = nlist->iinr;
1124     jindex           = nlist->jindex;
1125     jjnr             = nlist->jjnr;
1126     shiftidx         = nlist->shift;
1127     gid              = nlist->gid;
1128     shiftvec         = fr->shift_vec[0];
1129     fshift           = fr->fshift[0];
1130     facel            = _mm_set1_pd(fr->epsfac);
1131     charge           = mdatoms->chargeA;
1132     krf              = _mm_set1_pd(fr->ic->k_rf);
1133     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1134     crf              = _mm_set1_pd(fr->ic->c_rf);
1135
1136     /* Setup water-specific parameters */
1137     inr              = nlist->iinr[0];
1138     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1139     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1140     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1141
1142     jq1              = _mm_set1_pd(charge[inr+1]);
1143     jq2              = _mm_set1_pd(charge[inr+2]);
1144     jq3              = _mm_set1_pd(charge[inr+3]);
1145     qq11             = _mm_mul_pd(iq1,jq1);
1146     qq12             = _mm_mul_pd(iq1,jq2);
1147     qq13             = _mm_mul_pd(iq1,jq3);
1148     qq21             = _mm_mul_pd(iq2,jq1);
1149     qq22             = _mm_mul_pd(iq2,jq2);
1150     qq23             = _mm_mul_pd(iq2,jq3);
1151     qq31             = _mm_mul_pd(iq3,jq1);
1152     qq32             = _mm_mul_pd(iq3,jq2);
1153     qq33             = _mm_mul_pd(iq3,jq3);
1154
1155     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1156     rcutoff_scalar   = fr->rcoulomb;
1157     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1158     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1159
1160     /* Avoid stupid compiler warnings */
1161     jnrA = jnrB = 0;
1162     j_coord_offsetA = 0;
1163     j_coord_offsetB = 0;
1164
1165     outeriter        = 0;
1166     inneriter        = 0;
1167
1168     /* Start outer loop over neighborlists */
1169     for(iidx=0; iidx<nri; iidx++)
1170     {
1171         /* Load shift vector for this list */
1172         i_shift_offset   = DIM*shiftidx[iidx];
1173
1174         /* Load limits for loop over neighbors */
1175         j_index_start    = jindex[iidx];
1176         j_index_end      = jindex[iidx+1];
1177
1178         /* Get outer coordinate index */
1179         inr              = iinr[iidx];
1180         i_coord_offset   = DIM*inr;
1181
1182         /* Load i particle coords and add shift vector */
1183         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
1184                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1185
1186         fix1             = _mm_setzero_pd();
1187         fiy1             = _mm_setzero_pd();
1188         fiz1             = _mm_setzero_pd();
1189         fix2             = _mm_setzero_pd();
1190         fiy2             = _mm_setzero_pd();
1191         fiz2             = _mm_setzero_pd();
1192         fix3             = _mm_setzero_pd();
1193         fiy3             = _mm_setzero_pd();
1194         fiz3             = _mm_setzero_pd();
1195
1196         /* Start inner kernel loop */
1197         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1198         {
1199
1200             /* Get j neighbor index, and coordinate index */
1201             jnrA             = jjnr[jidx];
1202             jnrB             = jjnr[jidx+1];
1203             j_coord_offsetA  = DIM*jnrA;
1204             j_coord_offsetB  = DIM*jnrB;
1205
1206             /* load j atom coordinates */
1207             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1208                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1209
1210             /* Calculate displacement vector */
1211             dx11             = _mm_sub_pd(ix1,jx1);
1212             dy11             = _mm_sub_pd(iy1,jy1);
1213             dz11             = _mm_sub_pd(iz1,jz1);
1214             dx12             = _mm_sub_pd(ix1,jx2);
1215             dy12             = _mm_sub_pd(iy1,jy2);
1216             dz12             = _mm_sub_pd(iz1,jz2);
1217             dx13             = _mm_sub_pd(ix1,jx3);
1218             dy13             = _mm_sub_pd(iy1,jy3);
1219             dz13             = _mm_sub_pd(iz1,jz3);
1220             dx21             = _mm_sub_pd(ix2,jx1);
1221             dy21             = _mm_sub_pd(iy2,jy1);
1222             dz21             = _mm_sub_pd(iz2,jz1);
1223             dx22             = _mm_sub_pd(ix2,jx2);
1224             dy22             = _mm_sub_pd(iy2,jy2);
1225             dz22             = _mm_sub_pd(iz2,jz2);
1226             dx23             = _mm_sub_pd(ix2,jx3);
1227             dy23             = _mm_sub_pd(iy2,jy3);
1228             dz23             = _mm_sub_pd(iz2,jz3);
1229             dx31             = _mm_sub_pd(ix3,jx1);
1230             dy31             = _mm_sub_pd(iy3,jy1);
1231             dz31             = _mm_sub_pd(iz3,jz1);
1232             dx32             = _mm_sub_pd(ix3,jx2);
1233             dy32             = _mm_sub_pd(iy3,jy2);
1234             dz32             = _mm_sub_pd(iz3,jz2);
1235             dx33             = _mm_sub_pd(ix3,jx3);
1236             dy33             = _mm_sub_pd(iy3,jy3);
1237             dz33             = _mm_sub_pd(iz3,jz3);
1238
1239             /* Calculate squared distance and things based on it */
1240             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1241             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1242             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1243             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1244             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1245             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1246             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1247             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1248             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1249
1250             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1251             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1252             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1253             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1254             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1255             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1256             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1257             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1258             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1259
1260             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1261             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1262             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1263             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1264             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1265             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1266             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1267             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1268             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1269
1270             fjx1             = _mm_setzero_pd();
1271             fjy1             = _mm_setzero_pd();
1272             fjz1             = _mm_setzero_pd();
1273             fjx2             = _mm_setzero_pd();
1274             fjy2             = _mm_setzero_pd();
1275             fjz2             = _mm_setzero_pd();
1276             fjx3             = _mm_setzero_pd();
1277             fjy3             = _mm_setzero_pd();
1278             fjz3             = _mm_setzero_pd();
1279
1280             /**************************
1281              * CALCULATE INTERACTIONS *
1282              **************************/
1283
1284             if (gmx_mm_any_lt(rsq11,rcutoff2))
1285             {
1286
1287             /* REACTION-FIELD ELECTROSTATICS */
1288             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1289
1290             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1291
1292             fscal            = felec;
1293
1294             fscal            = _mm_and_pd(fscal,cutoff_mask);
1295
1296             /* Calculate temporary vectorial force */
1297             tx               = _mm_mul_pd(fscal,dx11);
1298             ty               = _mm_mul_pd(fscal,dy11);
1299             tz               = _mm_mul_pd(fscal,dz11);
1300
1301             /* Update vectorial force */
1302             fix1             = _mm_add_pd(fix1,tx);
1303             fiy1             = _mm_add_pd(fiy1,ty);
1304             fiz1             = _mm_add_pd(fiz1,tz);
1305
1306             fjx1             = _mm_add_pd(fjx1,tx);
1307             fjy1             = _mm_add_pd(fjy1,ty);
1308             fjz1             = _mm_add_pd(fjz1,tz);
1309
1310             }
1311
1312             /**************************
1313              * CALCULATE INTERACTIONS *
1314              **************************/
1315
1316             if (gmx_mm_any_lt(rsq12,rcutoff2))
1317             {
1318
1319             /* REACTION-FIELD ELECTROSTATICS */
1320             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1321
1322             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1323
1324             fscal            = felec;
1325
1326             fscal            = _mm_and_pd(fscal,cutoff_mask);
1327
1328             /* Calculate temporary vectorial force */
1329             tx               = _mm_mul_pd(fscal,dx12);
1330             ty               = _mm_mul_pd(fscal,dy12);
1331             tz               = _mm_mul_pd(fscal,dz12);
1332
1333             /* Update vectorial force */
1334             fix1             = _mm_add_pd(fix1,tx);
1335             fiy1             = _mm_add_pd(fiy1,ty);
1336             fiz1             = _mm_add_pd(fiz1,tz);
1337
1338             fjx2             = _mm_add_pd(fjx2,tx);
1339             fjy2             = _mm_add_pd(fjy2,ty);
1340             fjz2             = _mm_add_pd(fjz2,tz);
1341
1342             }
1343
1344             /**************************
1345              * CALCULATE INTERACTIONS *
1346              **************************/
1347
1348             if (gmx_mm_any_lt(rsq13,rcutoff2))
1349             {
1350
1351             /* REACTION-FIELD ELECTROSTATICS */
1352             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
1353
1354             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1355
1356             fscal            = felec;
1357
1358             fscal            = _mm_and_pd(fscal,cutoff_mask);
1359
1360             /* Calculate temporary vectorial force */
1361             tx               = _mm_mul_pd(fscal,dx13);
1362             ty               = _mm_mul_pd(fscal,dy13);
1363             tz               = _mm_mul_pd(fscal,dz13);
1364
1365             /* Update vectorial force */
1366             fix1             = _mm_add_pd(fix1,tx);
1367             fiy1             = _mm_add_pd(fiy1,ty);
1368             fiz1             = _mm_add_pd(fiz1,tz);
1369
1370             fjx3             = _mm_add_pd(fjx3,tx);
1371             fjy3             = _mm_add_pd(fjy3,ty);
1372             fjz3             = _mm_add_pd(fjz3,tz);
1373
1374             }
1375
1376             /**************************
1377              * CALCULATE INTERACTIONS *
1378              **************************/
1379
1380             if (gmx_mm_any_lt(rsq21,rcutoff2))
1381             {
1382
1383             /* REACTION-FIELD ELECTROSTATICS */
1384             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1385
1386             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1387
1388             fscal            = felec;
1389
1390             fscal            = _mm_and_pd(fscal,cutoff_mask);
1391
1392             /* Calculate temporary vectorial force */
1393             tx               = _mm_mul_pd(fscal,dx21);
1394             ty               = _mm_mul_pd(fscal,dy21);
1395             tz               = _mm_mul_pd(fscal,dz21);
1396
1397             /* Update vectorial force */
1398             fix2             = _mm_add_pd(fix2,tx);
1399             fiy2             = _mm_add_pd(fiy2,ty);
1400             fiz2             = _mm_add_pd(fiz2,tz);
1401
1402             fjx1             = _mm_add_pd(fjx1,tx);
1403             fjy1             = _mm_add_pd(fjy1,ty);
1404             fjz1             = _mm_add_pd(fjz1,tz);
1405
1406             }
1407
1408             /**************************
1409              * CALCULATE INTERACTIONS *
1410              **************************/
1411
1412             if (gmx_mm_any_lt(rsq22,rcutoff2))
1413             {
1414
1415             /* REACTION-FIELD ELECTROSTATICS */
1416             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1417
1418             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1419
1420             fscal            = felec;
1421
1422             fscal            = _mm_and_pd(fscal,cutoff_mask);
1423
1424             /* Calculate temporary vectorial force */
1425             tx               = _mm_mul_pd(fscal,dx22);
1426             ty               = _mm_mul_pd(fscal,dy22);
1427             tz               = _mm_mul_pd(fscal,dz22);
1428
1429             /* Update vectorial force */
1430             fix2             = _mm_add_pd(fix2,tx);
1431             fiy2             = _mm_add_pd(fiy2,ty);
1432             fiz2             = _mm_add_pd(fiz2,tz);
1433
1434             fjx2             = _mm_add_pd(fjx2,tx);
1435             fjy2             = _mm_add_pd(fjy2,ty);
1436             fjz2             = _mm_add_pd(fjz2,tz);
1437
1438             }
1439
1440             /**************************
1441              * CALCULATE INTERACTIONS *
1442              **************************/
1443
1444             if (gmx_mm_any_lt(rsq23,rcutoff2))
1445             {
1446
1447             /* REACTION-FIELD ELECTROSTATICS */
1448             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1449
1450             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1451
1452             fscal            = felec;
1453
1454             fscal            = _mm_and_pd(fscal,cutoff_mask);
1455
1456             /* Calculate temporary vectorial force */
1457             tx               = _mm_mul_pd(fscal,dx23);
1458             ty               = _mm_mul_pd(fscal,dy23);
1459             tz               = _mm_mul_pd(fscal,dz23);
1460
1461             /* Update vectorial force */
1462             fix2             = _mm_add_pd(fix2,tx);
1463             fiy2             = _mm_add_pd(fiy2,ty);
1464             fiz2             = _mm_add_pd(fiz2,tz);
1465
1466             fjx3             = _mm_add_pd(fjx3,tx);
1467             fjy3             = _mm_add_pd(fjy3,ty);
1468             fjz3             = _mm_add_pd(fjz3,tz);
1469
1470             }
1471
1472             /**************************
1473              * CALCULATE INTERACTIONS *
1474              **************************/
1475
1476             if (gmx_mm_any_lt(rsq31,rcutoff2))
1477             {
1478
1479             /* REACTION-FIELD ELECTROSTATICS */
1480             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1481
1482             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1483
1484             fscal            = felec;
1485
1486             fscal            = _mm_and_pd(fscal,cutoff_mask);
1487
1488             /* Calculate temporary vectorial force */
1489             tx               = _mm_mul_pd(fscal,dx31);
1490             ty               = _mm_mul_pd(fscal,dy31);
1491             tz               = _mm_mul_pd(fscal,dz31);
1492
1493             /* Update vectorial force */
1494             fix3             = _mm_add_pd(fix3,tx);
1495             fiy3             = _mm_add_pd(fiy3,ty);
1496             fiz3             = _mm_add_pd(fiz3,tz);
1497
1498             fjx1             = _mm_add_pd(fjx1,tx);
1499             fjy1             = _mm_add_pd(fjy1,ty);
1500             fjz1             = _mm_add_pd(fjz1,tz);
1501
1502             }
1503
1504             /**************************
1505              * CALCULATE INTERACTIONS *
1506              **************************/
1507
1508             if (gmx_mm_any_lt(rsq32,rcutoff2))
1509             {
1510
1511             /* REACTION-FIELD ELECTROSTATICS */
1512             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1513
1514             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1515
1516             fscal            = felec;
1517
1518             fscal            = _mm_and_pd(fscal,cutoff_mask);
1519
1520             /* Calculate temporary vectorial force */
1521             tx               = _mm_mul_pd(fscal,dx32);
1522             ty               = _mm_mul_pd(fscal,dy32);
1523             tz               = _mm_mul_pd(fscal,dz32);
1524
1525             /* Update vectorial force */
1526             fix3             = _mm_add_pd(fix3,tx);
1527             fiy3             = _mm_add_pd(fiy3,ty);
1528             fiz3             = _mm_add_pd(fiz3,tz);
1529
1530             fjx2             = _mm_add_pd(fjx2,tx);
1531             fjy2             = _mm_add_pd(fjy2,ty);
1532             fjz2             = _mm_add_pd(fjz2,tz);
1533
1534             }
1535
1536             /**************************
1537              * CALCULATE INTERACTIONS *
1538              **************************/
1539
1540             if (gmx_mm_any_lt(rsq33,rcutoff2))
1541             {
1542
1543             /* REACTION-FIELD ELECTROSTATICS */
1544             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1545
1546             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1547
1548             fscal            = felec;
1549
1550             fscal            = _mm_and_pd(fscal,cutoff_mask);
1551
1552             /* Calculate temporary vectorial force */
1553             tx               = _mm_mul_pd(fscal,dx33);
1554             ty               = _mm_mul_pd(fscal,dy33);
1555             tz               = _mm_mul_pd(fscal,dz33);
1556
1557             /* Update vectorial force */
1558             fix3             = _mm_add_pd(fix3,tx);
1559             fiy3             = _mm_add_pd(fiy3,ty);
1560             fiz3             = _mm_add_pd(fiz3,tz);
1561
1562             fjx3             = _mm_add_pd(fjx3,tx);
1563             fjy3             = _mm_add_pd(fjy3,ty);
1564             fjz3             = _mm_add_pd(fjz3,tz);
1565
1566             }
1567
1568             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);
1569
1570             /* Inner loop uses 270 flops */
1571         }
1572
1573         if(jidx<j_index_end)
1574         {
1575
1576             jnrA             = jjnr[jidx];
1577             j_coord_offsetA  = DIM*jnrA;
1578
1579             /* load j atom coordinates */
1580             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
1581                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1582
1583             /* Calculate displacement vector */
1584             dx11             = _mm_sub_pd(ix1,jx1);
1585             dy11             = _mm_sub_pd(iy1,jy1);
1586             dz11             = _mm_sub_pd(iz1,jz1);
1587             dx12             = _mm_sub_pd(ix1,jx2);
1588             dy12             = _mm_sub_pd(iy1,jy2);
1589             dz12             = _mm_sub_pd(iz1,jz2);
1590             dx13             = _mm_sub_pd(ix1,jx3);
1591             dy13             = _mm_sub_pd(iy1,jy3);
1592             dz13             = _mm_sub_pd(iz1,jz3);
1593             dx21             = _mm_sub_pd(ix2,jx1);
1594             dy21             = _mm_sub_pd(iy2,jy1);
1595             dz21             = _mm_sub_pd(iz2,jz1);
1596             dx22             = _mm_sub_pd(ix2,jx2);
1597             dy22             = _mm_sub_pd(iy2,jy2);
1598             dz22             = _mm_sub_pd(iz2,jz2);
1599             dx23             = _mm_sub_pd(ix2,jx3);
1600             dy23             = _mm_sub_pd(iy2,jy3);
1601             dz23             = _mm_sub_pd(iz2,jz3);
1602             dx31             = _mm_sub_pd(ix3,jx1);
1603             dy31             = _mm_sub_pd(iy3,jy1);
1604             dz31             = _mm_sub_pd(iz3,jz1);
1605             dx32             = _mm_sub_pd(ix3,jx2);
1606             dy32             = _mm_sub_pd(iy3,jy2);
1607             dz32             = _mm_sub_pd(iz3,jz2);
1608             dx33             = _mm_sub_pd(ix3,jx3);
1609             dy33             = _mm_sub_pd(iy3,jy3);
1610             dz33             = _mm_sub_pd(iz3,jz3);
1611
1612             /* Calculate squared distance and things based on it */
1613             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1614             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1615             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1616             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1617             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1618             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1619             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1620             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1621             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1622
1623             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1624             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1625             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1626             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1627             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1628             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1629             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1630             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1631             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1632
1633             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1634             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1635             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1636             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1637             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1638             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1639             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1640             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1641             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1642
1643             fjx1             = _mm_setzero_pd();
1644             fjy1             = _mm_setzero_pd();
1645             fjz1             = _mm_setzero_pd();
1646             fjx2             = _mm_setzero_pd();
1647             fjy2             = _mm_setzero_pd();
1648             fjz2             = _mm_setzero_pd();
1649             fjx3             = _mm_setzero_pd();
1650             fjy3             = _mm_setzero_pd();
1651             fjz3             = _mm_setzero_pd();
1652
1653             /**************************
1654              * CALCULATE INTERACTIONS *
1655              **************************/
1656
1657             if (gmx_mm_any_lt(rsq11,rcutoff2))
1658             {
1659
1660             /* REACTION-FIELD ELECTROSTATICS */
1661             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1662
1663             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1664
1665             fscal            = felec;
1666
1667             fscal            = _mm_and_pd(fscal,cutoff_mask);
1668
1669             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1670
1671             /* Calculate temporary vectorial force */
1672             tx               = _mm_mul_pd(fscal,dx11);
1673             ty               = _mm_mul_pd(fscal,dy11);
1674             tz               = _mm_mul_pd(fscal,dz11);
1675
1676             /* Update vectorial force */
1677             fix1             = _mm_add_pd(fix1,tx);
1678             fiy1             = _mm_add_pd(fiy1,ty);
1679             fiz1             = _mm_add_pd(fiz1,tz);
1680
1681             fjx1             = _mm_add_pd(fjx1,tx);
1682             fjy1             = _mm_add_pd(fjy1,ty);
1683             fjz1             = _mm_add_pd(fjz1,tz);
1684
1685             }
1686
1687             /**************************
1688              * CALCULATE INTERACTIONS *
1689              **************************/
1690
1691             if (gmx_mm_any_lt(rsq12,rcutoff2))
1692             {
1693
1694             /* REACTION-FIELD ELECTROSTATICS */
1695             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1696
1697             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1698
1699             fscal            = felec;
1700
1701             fscal            = _mm_and_pd(fscal,cutoff_mask);
1702
1703             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1704
1705             /* Calculate temporary vectorial force */
1706             tx               = _mm_mul_pd(fscal,dx12);
1707             ty               = _mm_mul_pd(fscal,dy12);
1708             tz               = _mm_mul_pd(fscal,dz12);
1709
1710             /* Update vectorial force */
1711             fix1             = _mm_add_pd(fix1,tx);
1712             fiy1             = _mm_add_pd(fiy1,ty);
1713             fiz1             = _mm_add_pd(fiz1,tz);
1714
1715             fjx2             = _mm_add_pd(fjx2,tx);
1716             fjy2             = _mm_add_pd(fjy2,ty);
1717             fjz2             = _mm_add_pd(fjz2,tz);
1718
1719             }
1720
1721             /**************************
1722              * CALCULATE INTERACTIONS *
1723              **************************/
1724
1725             if (gmx_mm_any_lt(rsq13,rcutoff2))
1726             {
1727
1728             /* REACTION-FIELD ELECTROSTATICS */
1729             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
1730
1731             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1732
1733             fscal            = felec;
1734
1735             fscal            = _mm_and_pd(fscal,cutoff_mask);
1736
1737             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1738
1739             /* Calculate temporary vectorial force */
1740             tx               = _mm_mul_pd(fscal,dx13);
1741             ty               = _mm_mul_pd(fscal,dy13);
1742             tz               = _mm_mul_pd(fscal,dz13);
1743
1744             /* Update vectorial force */
1745             fix1             = _mm_add_pd(fix1,tx);
1746             fiy1             = _mm_add_pd(fiy1,ty);
1747             fiz1             = _mm_add_pd(fiz1,tz);
1748
1749             fjx3             = _mm_add_pd(fjx3,tx);
1750             fjy3             = _mm_add_pd(fjy3,ty);
1751             fjz3             = _mm_add_pd(fjz3,tz);
1752
1753             }
1754
1755             /**************************
1756              * CALCULATE INTERACTIONS *
1757              **************************/
1758
1759             if (gmx_mm_any_lt(rsq21,rcutoff2))
1760             {
1761
1762             /* REACTION-FIELD ELECTROSTATICS */
1763             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1764
1765             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1766
1767             fscal            = felec;
1768
1769             fscal            = _mm_and_pd(fscal,cutoff_mask);
1770
1771             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1772
1773             /* Calculate temporary vectorial force */
1774             tx               = _mm_mul_pd(fscal,dx21);
1775             ty               = _mm_mul_pd(fscal,dy21);
1776             tz               = _mm_mul_pd(fscal,dz21);
1777
1778             /* Update vectorial force */
1779             fix2             = _mm_add_pd(fix2,tx);
1780             fiy2             = _mm_add_pd(fiy2,ty);
1781             fiz2             = _mm_add_pd(fiz2,tz);
1782
1783             fjx1             = _mm_add_pd(fjx1,tx);
1784             fjy1             = _mm_add_pd(fjy1,ty);
1785             fjz1             = _mm_add_pd(fjz1,tz);
1786
1787             }
1788
1789             /**************************
1790              * CALCULATE INTERACTIONS *
1791              **************************/
1792
1793             if (gmx_mm_any_lt(rsq22,rcutoff2))
1794             {
1795
1796             /* REACTION-FIELD ELECTROSTATICS */
1797             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1798
1799             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1800
1801             fscal            = felec;
1802
1803             fscal            = _mm_and_pd(fscal,cutoff_mask);
1804
1805             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1806
1807             /* Calculate temporary vectorial force */
1808             tx               = _mm_mul_pd(fscal,dx22);
1809             ty               = _mm_mul_pd(fscal,dy22);
1810             tz               = _mm_mul_pd(fscal,dz22);
1811
1812             /* Update vectorial force */
1813             fix2             = _mm_add_pd(fix2,tx);
1814             fiy2             = _mm_add_pd(fiy2,ty);
1815             fiz2             = _mm_add_pd(fiz2,tz);
1816
1817             fjx2             = _mm_add_pd(fjx2,tx);
1818             fjy2             = _mm_add_pd(fjy2,ty);
1819             fjz2             = _mm_add_pd(fjz2,tz);
1820
1821             }
1822
1823             /**************************
1824              * CALCULATE INTERACTIONS *
1825              **************************/
1826
1827             if (gmx_mm_any_lt(rsq23,rcutoff2))
1828             {
1829
1830             /* REACTION-FIELD ELECTROSTATICS */
1831             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1832
1833             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1834
1835             fscal            = felec;
1836
1837             fscal            = _mm_and_pd(fscal,cutoff_mask);
1838
1839             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1840
1841             /* Calculate temporary vectorial force */
1842             tx               = _mm_mul_pd(fscal,dx23);
1843             ty               = _mm_mul_pd(fscal,dy23);
1844             tz               = _mm_mul_pd(fscal,dz23);
1845
1846             /* Update vectorial force */
1847             fix2             = _mm_add_pd(fix2,tx);
1848             fiy2             = _mm_add_pd(fiy2,ty);
1849             fiz2             = _mm_add_pd(fiz2,tz);
1850
1851             fjx3             = _mm_add_pd(fjx3,tx);
1852             fjy3             = _mm_add_pd(fjy3,ty);
1853             fjz3             = _mm_add_pd(fjz3,tz);
1854
1855             }
1856
1857             /**************************
1858              * CALCULATE INTERACTIONS *
1859              **************************/
1860
1861             if (gmx_mm_any_lt(rsq31,rcutoff2))
1862             {
1863
1864             /* REACTION-FIELD ELECTROSTATICS */
1865             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1866
1867             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1868
1869             fscal            = felec;
1870
1871             fscal            = _mm_and_pd(fscal,cutoff_mask);
1872
1873             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1874
1875             /* Calculate temporary vectorial force */
1876             tx               = _mm_mul_pd(fscal,dx31);
1877             ty               = _mm_mul_pd(fscal,dy31);
1878             tz               = _mm_mul_pd(fscal,dz31);
1879
1880             /* Update vectorial force */
1881             fix3             = _mm_add_pd(fix3,tx);
1882             fiy3             = _mm_add_pd(fiy3,ty);
1883             fiz3             = _mm_add_pd(fiz3,tz);
1884
1885             fjx1             = _mm_add_pd(fjx1,tx);
1886             fjy1             = _mm_add_pd(fjy1,ty);
1887             fjz1             = _mm_add_pd(fjz1,tz);
1888
1889             }
1890
1891             /**************************
1892              * CALCULATE INTERACTIONS *
1893              **************************/
1894
1895             if (gmx_mm_any_lt(rsq32,rcutoff2))
1896             {
1897
1898             /* REACTION-FIELD ELECTROSTATICS */
1899             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1900
1901             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1902
1903             fscal            = felec;
1904
1905             fscal            = _mm_and_pd(fscal,cutoff_mask);
1906
1907             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1908
1909             /* Calculate temporary vectorial force */
1910             tx               = _mm_mul_pd(fscal,dx32);
1911             ty               = _mm_mul_pd(fscal,dy32);
1912             tz               = _mm_mul_pd(fscal,dz32);
1913
1914             /* Update vectorial force */
1915             fix3             = _mm_add_pd(fix3,tx);
1916             fiy3             = _mm_add_pd(fiy3,ty);
1917             fiz3             = _mm_add_pd(fiz3,tz);
1918
1919             fjx2             = _mm_add_pd(fjx2,tx);
1920             fjy2             = _mm_add_pd(fjy2,ty);
1921             fjz2             = _mm_add_pd(fjz2,tz);
1922
1923             }
1924
1925             /**************************
1926              * CALCULATE INTERACTIONS *
1927              **************************/
1928
1929             if (gmx_mm_any_lt(rsq33,rcutoff2))
1930             {
1931
1932             /* REACTION-FIELD ELECTROSTATICS */
1933             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1934
1935             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1936
1937             fscal            = felec;
1938
1939             fscal            = _mm_and_pd(fscal,cutoff_mask);
1940
1941             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1942
1943             /* Calculate temporary vectorial force */
1944             tx               = _mm_mul_pd(fscal,dx33);
1945             ty               = _mm_mul_pd(fscal,dy33);
1946             tz               = _mm_mul_pd(fscal,dz33);
1947
1948             /* Update vectorial force */
1949             fix3             = _mm_add_pd(fix3,tx);
1950             fiy3             = _mm_add_pd(fiy3,ty);
1951             fiz3             = _mm_add_pd(fiz3,tz);
1952
1953             fjx3             = _mm_add_pd(fjx3,tx);
1954             fjy3             = _mm_add_pd(fjy3,ty);
1955             fjz3             = _mm_add_pd(fjz3,tz);
1956
1957             }
1958
1959             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1960
1961             /* Inner loop uses 270 flops */
1962         }
1963
1964         /* End of innermost loop */
1965
1966         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1967                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1968
1969         /* Increment number of inner iterations */
1970         inneriter                  += j_index_end - j_index_start;
1971
1972         /* Outer loop uses 18 flops */
1973     }
1974
1975     /* Increment number of outer iterations */
1976     outeriter        += nri;
1977
1978     /* Update outer/inner flops */
1979
1980     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*270);
1981 }