src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecRFCut_VdwNone_GeomW4W4_avx_128_fma_double.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_128_fma_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_avx_128_fma_double.h"
  34 #include "kernelutil_x86_avx_128_fma_double.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwNone_GeomW4W4_VF_avx_128_fma_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_avx_128_fma_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_macc_pd(krf,rsq11,rinv11),crf));
 266             felec            = _mm_mul_pd(qq11,_mm_msub_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             /* Update vectorial force */
 279             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 280             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 281             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 282
 283             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 284             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 285             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 286
 287             }
 288
 289             /**************************
 290              * CALCULATE INTERACTIONS *
 291              **************************/
 292
 293             if (gmx_mm_any_lt(rsq12,rcutoff2))
 294             {
 295
 296             /* REACTION-FIELD ELECTROSTATICS */
 297             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 298             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 299
 300             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 301
 302             /* Update potential sum for this i atom from the interaction with this j atom. */
 303             velec            = _mm_and_pd(velec,cutoff_mask);
 304             velecsum         = _mm_add_pd(velecsum,velec);
 305
 306             fscal            = felec;
 307
 308             fscal            = _mm_and_pd(fscal,cutoff_mask);
 309
 310             /* Update vectorial force */
 311             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 312             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 313             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 314
 315             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 316             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 317             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 318
 319             }
 320
 321             /**************************
 322              * CALCULATE INTERACTIONS *
 323              **************************/
 324
 325             if (gmx_mm_any_lt(rsq13,rcutoff2))
 326             {
 327
 328             /* REACTION-FIELD ELECTROSTATICS */
 329             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 330             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 331
 332             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 333
 334             /* Update potential sum for this i atom from the interaction with this j atom. */
 335             velec            = _mm_and_pd(velec,cutoff_mask);
 336             velecsum         = _mm_add_pd(velecsum,velec);
 337
 338             fscal            = felec;
 339
 340             fscal            = _mm_and_pd(fscal,cutoff_mask);
 341
 342             /* Update vectorial force */
 343             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 344             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 345             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 346
 347             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 348             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 349             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 350
 351             }
 352
 353             /**************************
 354              * CALCULATE INTERACTIONS *
 355              **************************/
 356
 357             if (gmx_mm_any_lt(rsq21,rcutoff2))
 358             {
 359
 360             /* REACTION-FIELD ELECTROSTATICS */
 361             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 362             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 363
 364             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 365
 366             /* Update potential sum for this i atom from the interaction with this j atom. */
 367             velec            = _mm_and_pd(velec,cutoff_mask);
 368             velecsum         = _mm_add_pd(velecsum,velec);
 369
 370             fscal            = felec;
 371
 372             fscal            = _mm_and_pd(fscal,cutoff_mask);
 373
 374             /* Update vectorial force */
 375             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 376             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 377             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 378
 379             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 380             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 381             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 382
 383             }
 384
 385             /**************************
 386              * CALCULATE INTERACTIONS *
 387              **************************/
 388
 389             if (gmx_mm_any_lt(rsq22,rcutoff2))
 390             {
 391
 392             /* REACTION-FIELD ELECTROSTATICS */
 393             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 394             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 395
 396             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 397
 398             /* Update potential sum for this i atom from the interaction with this j atom. */
 399             velec            = _mm_and_pd(velec,cutoff_mask);
 400             velecsum         = _mm_add_pd(velecsum,velec);
 401
 402             fscal            = felec;
 403
 404             fscal            = _mm_and_pd(fscal,cutoff_mask);
 405
 406             /* Update vectorial force */
 407             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 408             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 409             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 410
 411             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 412             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 413             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 414
 415             }
 416
 417             /**************************
 418              * CALCULATE INTERACTIONS *
 419              **************************/
 420
 421             if (gmx_mm_any_lt(rsq23,rcutoff2))
 422             {
 423
 424             /* REACTION-FIELD ELECTROSTATICS */
 425             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 426             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 427
 428             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 429
 430             /* Update potential sum for this i atom from the interaction with this j atom. */
 431             velec            = _mm_and_pd(velec,cutoff_mask);
 432             velecsum         = _mm_add_pd(velecsum,velec);
 433
 434             fscal            = felec;
 435
 436             fscal            = _mm_and_pd(fscal,cutoff_mask);
 437
 438             /* Update vectorial force */
 439             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 440             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 441             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 442
 443             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 444             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 445             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 446
 447             }
 448
 449             /**************************
 450              * CALCULATE INTERACTIONS *
 451              **************************/
 452
 453             if (gmx_mm_any_lt(rsq31,rcutoff2))
 454             {
 455
 456             /* REACTION-FIELD ELECTROSTATICS */
 457             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
 458             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
 459
 460             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
 461
 462             /* Update potential sum for this i atom from the interaction with this j atom. */
 463             velec            = _mm_and_pd(velec,cutoff_mask);
 464             velecsum         = _mm_add_pd(velecsum,velec);
 465
 466             fscal            = felec;
 467
 468             fscal            = _mm_and_pd(fscal,cutoff_mask);
 469
 470             /* Update vectorial force */
 471             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 472             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 473             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 474
 475             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 476             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 477             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 478
 479             }
 480
 481             /**************************
 482              * CALCULATE INTERACTIONS *
 483              **************************/
 484
 485             if (gmx_mm_any_lt(rsq32,rcutoff2))
 486             {
 487
 488             /* REACTION-FIELD ELECTROSTATICS */
 489             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
 490             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
 491
 492             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
 493
 494             /* Update potential sum for this i atom from the interaction with this j atom. */
 495             velec            = _mm_and_pd(velec,cutoff_mask);
 496             velecsum         = _mm_add_pd(velecsum,velec);
 497
 498             fscal            = felec;
 499
 500             fscal            = _mm_and_pd(fscal,cutoff_mask);
 501
 502             /* Update vectorial force */
 503             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 504             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 505             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 506
 507             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 508             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 509             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 510
 511             }
 512
 513             /**************************
 514              * CALCULATE INTERACTIONS *
 515              **************************/
 516
 517             if (gmx_mm_any_lt(rsq33,rcutoff2))
 518             {
 519
 520             /* REACTION-FIELD ELECTROSTATICS */
 521             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
 522             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
 523
 524             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
 525
 526             /* Update potential sum for this i atom from the interaction with this j atom. */
 527             velec            = _mm_and_pd(velec,cutoff_mask);
 528             velecsum         = _mm_add_pd(velecsum,velec);
 529
 530             fscal            = felec;
 531
 532             fscal            = _mm_and_pd(fscal,cutoff_mask);
 533
 534             /* Update vectorial force */
 535             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 536             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 537             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 538
 539             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 540             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 541             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 542
 543             }
 544
 545             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);
 546
 547             /* Inner loop uses 351 flops */
 548         }
 549
 550         if(jidx<j_index_end)
 551         {
 552
 553             jnrA             = jjnr[jidx];
 554             j_coord_offsetA  = DIM*jnrA;
 555
 556             /* load j atom coordinates */
 557             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
 558                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 559
 560             /* Calculate displacement vector */
 561             dx11             = _mm_sub_pd(ix1,jx1);
 562             dy11             = _mm_sub_pd(iy1,jy1);
 563             dz11             = _mm_sub_pd(iz1,jz1);
 564             dx12             = _mm_sub_pd(ix1,jx2);
 565             dy12             = _mm_sub_pd(iy1,jy2);
 566             dz12             = _mm_sub_pd(iz1,jz2);
 567             dx13             = _mm_sub_pd(ix1,jx3);
 568             dy13             = _mm_sub_pd(iy1,jy3);
 569             dz13             = _mm_sub_pd(iz1,jz3);
 570             dx21             = _mm_sub_pd(ix2,jx1);
 571             dy21             = _mm_sub_pd(iy2,jy1);
 572             dz21             = _mm_sub_pd(iz2,jz1);
 573             dx22             = _mm_sub_pd(ix2,jx2);
 574             dy22             = _mm_sub_pd(iy2,jy2);
 575             dz22             = _mm_sub_pd(iz2,jz2);
 576             dx23             = _mm_sub_pd(ix2,jx3);
 577             dy23             = _mm_sub_pd(iy2,jy3);
 578             dz23             = _mm_sub_pd(iz2,jz3);
 579             dx31             = _mm_sub_pd(ix3,jx1);
 580             dy31             = _mm_sub_pd(iy3,jy1);
 581             dz31             = _mm_sub_pd(iz3,jz1);
 582             dx32             = _mm_sub_pd(ix3,jx2);
 583             dy32             = _mm_sub_pd(iy3,jy2);
 584             dz32             = _mm_sub_pd(iz3,jz2);
 585             dx33             = _mm_sub_pd(ix3,jx3);
 586             dy33             = _mm_sub_pd(iy3,jy3);
 587             dz33             = _mm_sub_pd(iz3,jz3);
 588
 589             /* Calculate squared distance and things based on it */
 590             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 591             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 592             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 593             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 594             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 595             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 596             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 597             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 598             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 599
 600             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 601             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 602             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 603             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 604             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 605             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 606             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 607             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 608             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 609
 610             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 611             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 612             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 613             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 614             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 615             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 616             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 617             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 618             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 619
 620             fjx1             = _mm_setzero_pd();
 621             fjy1             = _mm_setzero_pd();
 622             fjz1             = _mm_setzero_pd();
 623             fjx2             = _mm_setzero_pd();
 624             fjy2             = _mm_setzero_pd();
 625             fjz2             = _mm_setzero_pd();
 626             fjx3             = _mm_setzero_pd();
 627             fjy3             = _mm_setzero_pd();
 628             fjz3             = _mm_setzero_pd();
 629
 630             /**************************
 631              * CALCULATE INTERACTIONS *
 632              **************************/
 633
 634             if (gmx_mm_any_lt(rsq11,rcutoff2))
 635             {
 636
 637             /* REACTION-FIELD ELECTROSTATICS */
 638             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 639             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 640
 641             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 642
 643             /* Update potential sum for this i atom from the interaction with this j atom. */
 644             velec            = _mm_and_pd(velec,cutoff_mask);
 645             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 646             velecsum         = _mm_add_pd(velecsum,velec);
 647
 648             fscal            = felec;
 649
 650             fscal            = _mm_and_pd(fscal,cutoff_mask);
 651
 652             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 653
 654             /* Update vectorial force */
 655             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 656             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 657             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 658
 659             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 660             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 661             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 662
 663             }
 664
 665             /**************************
 666              * CALCULATE INTERACTIONS *
 667              **************************/
 668
 669             if (gmx_mm_any_lt(rsq12,rcutoff2))
 670             {
 671
 672             /* REACTION-FIELD ELECTROSTATICS */
 673             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 674             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 675
 676             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 677
 678             /* Update potential sum for this i atom from the interaction with this j atom. */
 679             velec            = _mm_and_pd(velec,cutoff_mask);
 680             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 681             velecsum         = _mm_add_pd(velecsum,velec);
 682
 683             fscal            = felec;
 684
 685             fscal            = _mm_and_pd(fscal,cutoff_mask);
 686
 687             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 688
 689             /* Update vectorial force */
 690             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 691             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 692             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 693
 694             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 695             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 696             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 697
 698             }
 699
 700             /**************************
 701              * CALCULATE INTERACTIONS *
 702              **************************/
 703
 704             if (gmx_mm_any_lt(rsq13,rcutoff2))
 705             {
 706
 707             /* REACTION-FIELD ELECTROSTATICS */
 708             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 709             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 710
 711             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 712
 713             /* Update potential sum for this i atom from the interaction with this j atom. */
 714             velec            = _mm_and_pd(velec,cutoff_mask);
 715             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 716             velecsum         = _mm_add_pd(velecsum,velec);
 717
 718             fscal            = felec;
 719
 720             fscal            = _mm_and_pd(fscal,cutoff_mask);
 721
 722             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 723
 724             /* Update vectorial force */
 725             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 726             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 727             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 728
 729             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 730             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 731             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 732
 733             }
 734
 735             /**************************
 736              * CALCULATE INTERACTIONS *
 737              **************************/
 738
 739             if (gmx_mm_any_lt(rsq21,rcutoff2))
 740             {
 741
 742             /* REACTION-FIELD ELECTROSTATICS */
 743             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 744             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 745
 746             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 747
 748             /* Update potential sum for this i atom from the interaction with this j atom. */
 749             velec            = _mm_and_pd(velec,cutoff_mask);
 750             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 751             velecsum         = _mm_add_pd(velecsum,velec);
 752
 753             fscal            = felec;
 754
 755             fscal            = _mm_and_pd(fscal,cutoff_mask);
 756
 757             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 758
 759             /* Update vectorial force */
 760             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 761             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 762             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 763
 764             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 765             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 766             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 767
 768             }
 769
 770             /**************************
 771              * CALCULATE INTERACTIONS *
 772              **************************/
 773
 774             if (gmx_mm_any_lt(rsq22,rcutoff2))
 775             {
 776
 777             /* REACTION-FIELD ELECTROSTATICS */
 778             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 779             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 780
 781             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 782
 783             /* Update potential sum for this i atom from the interaction with this j atom. */
 784             velec            = _mm_and_pd(velec,cutoff_mask);
 785             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 786             velecsum         = _mm_add_pd(velecsum,velec);
 787
 788             fscal            = felec;
 789
 790             fscal            = _mm_and_pd(fscal,cutoff_mask);
 791
 792             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 793
 794             /* Update vectorial force */
 795             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 796             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 797             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 798
 799             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 800             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 801             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 802
 803             }
 804
 805             /**************************
 806              * CALCULATE INTERACTIONS *
 807              **************************/
 808
 809             if (gmx_mm_any_lt(rsq23,rcutoff2))
 810             {
 811
 812             /* REACTION-FIELD ELECTROSTATICS */
 813             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 814             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 815
 816             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 817
 818             /* Update potential sum for this i atom from the interaction with this j atom. */
 819             velec            = _mm_and_pd(velec,cutoff_mask);
 820             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 821             velecsum         = _mm_add_pd(velecsum,velec);
 822
 823             fscal            = felec;
 824
 825             fscal            = _mm_and_pd(fscal,cutoff_mask);
 826
 827             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 828
 829             /* Update vectorial force */
 830             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 831             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 832             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 833
 834             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 835             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 836             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 837
 838             }
 839
 840             /**************************
 841              * CALCULATE INTERACTIONS *
 842              **************************/
 843
 844             if (gmx_mm_any_lt(rsq31,rcutoff2))
 845             {
 846
 847             /* REACTION-FIELD ELECTROSTATICS */
 848             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
 849             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
 850
 851             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
 852
 853             /* Update potential sum for this i atom from the interaction with this j atom. */
 854             velec            = _mm_and_pd(velec,cutoff_mask);
 855             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 856             velecsum         = _mm_add_pd(velecsum,velec);
 857
 858             fscal            = felec;
 859
 860             fscal            = _mm_and_pd(fscal,cutoff_mask);
 861
 862             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 863
 864             /* Update vectorial force */
 865             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 866             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 867             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 868
 869             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 870             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 871             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 872
 873             }
 874
 875             /**************************
 876              * CALCULATE INTERACTIONS *
 877              **************************/
 878
 879             if (gmx_mm_any_lt(rsq32,rcutoff2))
 880             {
 881
 882             /* REACTION-FIELD ELECTROSTATICS */
 883             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
 884             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
 885
 886             cutoff_mask      = _mm_cmplt_pd(rsq32,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             /* Update vectorial force */
 900             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 901             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 902             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 903
 904             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 905             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 906             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 907
 908             }
 909
 910             /**************************
 911              * CALCULATE INTERACTIONS *
 912              **************************/
 913
 914             if (gmx_mm_any_lt(rsq33,rcutoff2))
 915             {
 916
 917             /* REACTION-FIELD ELECTROSTATICS */
 918             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
 919             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
 920
 921             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
 922
 923             /* Update potential sum for this i atom from the interaction with this j atom. */
 924             velec            = _mm_and_pd(velec,cutoff_mask);
 925             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 926             velecsum         = _mm_add_pd(velecsum,velec);
 927
 928             fscal            = felec;
 929
 930             fscal            = _mm_and_pd(fscal,cutoff_mask);
 931
 932             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 933
 934             /* Update vectorial force */
 935             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 936             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 937             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 938
 939             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 940             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 941             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 942
 943             }
 944
 945             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 946
 947             /* Inner loop uses 351 flops */
 948         }
 949
 950         /* End of innermost loop */
 951
 952         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 953                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
 954
 955         ggid                        = gid[iidx];
 956         /* Update potential energies */
 957         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 958
 959         /* Increment number of inner iterations */
 960         inneriter                  += j_index_end - j_index_start;
 961
 962         /* Outer loop uses 19 flops */
 963     }
 964
 965     /* Increment number of outer iterations */
 966     outeriter        += nri;
 967
 968     /* Update outer/inner flops */
 969
 970     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*351);
 971 }
 972 /*
 973  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwNone_GeomW4W4_F_avx_128_fma_double
 974  * Electrostatics interaction: ReactionField
 975  * VdW interaction:            None
 976  * Geometry:                   Water4-Water4
 977  * Calculate force/pot:        Force
 978  */
 979 void
 980 nb_kernel_ElecRFCut_VdwNone_GeomW4W4_F_avx_128_fma_double
 981                     (t_nblist * gmx_restrict                nlist,
 982                      rvec * gmx_restrict                    xx,
 983                      rvec * gmx_restrict                    ff,
 984                      t_forcerec * gmx_restrict              fr,
 985                      t_mdatoms * gmx_restrict               mdatoms,
 986                      nb_kernel_data_t * gmx_restrict        kernel_data,
 987                      t_nrnb * gmx_restrict                  nrnb)
 988 {
 989     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 990      * just 0 for non-waters.
 991      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 992      * jnr indices corresponding to data put in the four positions in the SIMD register.
 993      */
 994     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 995     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 996     int              jnrA,jnrB;
 997     int              j_coord_offsetA,j_coord_offsetB;
 998     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 999     real             rcutoff_scalar;
1000     real             *shiftvec,*fshift,*x,*f;
1001     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1002     int              vdwioffset1;
1003     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1004     int              vdwioffset2;
1005     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1006     int              vdwioffset3;
1007     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1008     int              vdwjidx1A,vdwjidx1B;
1009     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1010     int              vdwjidx2A,vdwjidx2B;
1011     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1012     int              vdwjidx3A,vdwjidx3B;
1013     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1014     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1015     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1016     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1017     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1018     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1019     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1020     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1021     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1022     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1023     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1024     real             *charge;
1025     __m128d          dummy_mask,cutoff_mask;
1026     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1027     __m128d          one     = _mm_set1_pd(1.0);
1028     __m128d          two     = _mm_set1_pd(2.0);
1029     x                = xx[0];
1030     f                = ff[0];
1031
1032     nri              = nlist->nri;
1033     iinr             = nlist->iinr;
1034     jindex           = nlist->jindex;
1035     jjnr             = nlist->jjnr;
1036     shiftidx         = nlist->shift;
1037     gid              = nlist->gid;
1038     shiftvec         = fr->shift_vec[0];
1039     fshift           = fr->fshift[0];
1040     facel            = _mm_set1_pd(fr->epsfac);
1041     charge           = mdatoms->chargeA;
1042     krf              = _mm_set1_pd(fr->ic->k_rf);
1043     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1044     crf              = _mm_set1_pd(fr->ic->c_rf);
1045
1046     /* Setup water-specific parameters */
1047     inr              = nlist->iinr[0];
1048     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1049     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1050     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1051
1052     jq1              = _mm_set1_pd(charge[inr+1]);
1053     jq2              = _mm_set1_pd(charge[inr+2]);
1054     jq3              = _mm_set1_pd(charge[inr+3]);
1055     qq11             = _mm_mul_pd(iq1,jq1);
1056     qq12             = _mm_mul_pd(iq1,jq2);
1057     qq13             = _mm_mul_pd(iq1,jq3);
1058     qq21             = _mm_mul_pd(iq2,jq1);
1059     qq22             = _mm_mul_pd(iq2,jq2);
1060     qq23             = _mm_mul_pd(iq2,jq3);
1061     qq31             = _mm_mul_pd(iq3,jq1);
1062     qq32             = _mm_mul_pd(iq3,jq2);
1063     qq33             = _mm_mul_pd(iq3,jq3);
1064
1065     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1066     rcutoff_scalar   = fr->rcoulomb;
1067     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1068     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1069
1070     /* Avoid stupid compiler warnings */
1071     jnrA = jnrB = 0;
1072     j_coord_offsetA = 0;
1073     j_coord_offsetB = 0;
1074
1075     outeriter        = 0;
1076     inneriter        = 0;
1077
1078     /* Start outer loop over neighborlists */
1079     for(iidx=0; iidx<nri; iidx++)
1080     {
1081         /* Load shift vector for this list */
1082         i_shift_offset   = DIM*shiftidx[iidx];
1083
1084         /* Load limits for loop over neighbors */
1085         j_index_start    = jindex[iidx];
1086         j_index_end      = jindex[iidx+1];
1087
1088         /* Get outer coordinate index */
1089         inr              = iinr[iidx];
1090         i_coord_offset   = DIM*inr;
1091
1092         /* Load i particle coords and add shift vector */
1093         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
1094                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1095
1096         fix1             = _mm_setzero_pd();
1097         fiy1             = _mm_setzero_pd();
1098         fiz1             = _mm_setzero_pd();
1099         fix2             = _mm_setzero_pd();
1100         fiy2             = _mm_setzero_pd();
1101         fiz2             = _mm_setzero_pd();
1102         fix3             = _mm_setzero_pd();
1103         fiy3             = _mm_setzero_pd();
1104         fiz3             = _mm_setzero_pd();
1105
1106         /* Start inner kernel loop */
1107         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1108         {
1109
1110             /* Get j neighbor index, and coordinate index */
1111             jnrA             = jjnr[jidx];
1112             jnrB             = jjnr[jidx+1];
1113             j_coord_offsetA  = DIM*jnrA;
1114             j_coord_offsetB  = DIM*jnrB;
1115
1116             /* load j atom coordinates */
1117             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1118                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1119
1120             /* Calculate displacement vector */
1121             dx11             = _mm_sub_pd(ix1,jx1);
1122             dy11             = _mm_sub_pd(iy1,jy1);
1123             dz11             = _mm_sub_pd(iz1,jz1);
1124             dx12             = _mm_sub_pd(ix1,jx2);
1125             dy12             = _mm_sub_pd(iy1,jy2);
1126             dz12             = _mm_sub_pd(iz1,jz2);
1127             dx13             = _mm_sub_pd(ix1,jx3);
1128             dy13             = _mm_sub_pd(iy1,jy3);
1129             dz13             = _mm_sub_pd(iz1,jz3);
1130             dx21             = _mm_sub_pd(ix2,jx1);
1131             dy21             = _mm_sub_pd(iy2,jy1);
1132             dz21             = _mm_sub_pd(iz2,jz1);
1133             dx22             = _mm_sub_pd(ix2,jx2);
1134             dy22             = _mm_sub_pd(iy2,jy2);
1135             dz22             = _mm_sub_pd(iz2,jz2);
1136             dx23             = _mm_sub_pd(ix2,jx3);
1137             dy23             = _mm_sub_pd(iy2,jy3);
1138             dz23             = _mm_sub_pd(iz2,jz3);
1139             dx31             = _mm_sub_pd(ix3,jx1);
1140             dy31             = _mm_sub_pd(iy3,jy1);
1141             dz31             = _mm_sub_pd(iz3,jz1);
1142             dx32             = _mm_sub_pd(ix3,jx2);
1143             dy32             = _mm_sub_pd(iy3,jy2);
1144             dz32             = _mm_sub_pd(iz3,jz2);
1145             dx33             = _mm_sub_pd(ix3,jx3);
1146             dy33             = _mm_sub_pd(iy3,jy3);
1147             dz33             = _mm_sub_pd(iz3,jz3);
1148
1149             /* Calculate squared distance and things based on it */
1150             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1151             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1152             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1153             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1154             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1155             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1156             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1157             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1158             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1159
1160             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1161             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1162             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1163             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1164             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1165             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1166             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1167             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1168             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1169
1170             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1171             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1172             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1173             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1174             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1175             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1176             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1177             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1178             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1179
1180             fjx1             = _mm_setzero_pd();
1181             fjy1             = _mm_setzero_pd();
1182             fjz1             = _mm_setzero_pd();
1183             fjx2             = _mm_setzero_pd();
1184             fjy2             = _mm_setzero_pd();
1185             fjz2             = _mm_setzero_pd();
1186             fjx3             = _mm_setzero_pd();
1187             fjy3             = _mm_setzero_pd();
1188             fjz3             = _mm_setzero_pd();
1189
1190             /**************************
1191              * CALCULATE INTERACTIONS *
1192              **************************/
1193
1194             if (gmx_mm_any_lt(rsq11,rcutoff2))
1195             {
1196
1197             /* REACTION-FIELD ELECTROSTATICS */
1198             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1199
1200             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1201
1202             fscal            = felec;
1203
1204             fscal            = _mm_and_pd(fscal,cutoff_mask);
1205
1206             /* Update vectorial force */
1207             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1208             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1209             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1210
1211             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1212             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1213             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1214
1215             }
1216
1217             /**************************
1218              * CALCULATE INTERACTIONS *
1219              **************************/
1220
1221             if (gmx_mm_any_lt(rsq12,rcutoff2))
1222             {
1223
1224             /* REACTION-FIELD ELECTROSTATICS */
1225             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1226
1227             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1228
1229             fscal            = felec;
1230
1231             fscal            = _mm_and_pd(fscal,cutoff_mask);
1232
1233             /* Update vectorial force */
1234             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1235             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1236             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1237
1238             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1239             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1240             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1241
1242             }
1243
1244             /**************************
1245              * CALCULATE INTERACTIONS *
1246              **************************/
1247
1248             if (gmx_mm_any_lt(rsq13,rcutoff2))
1249             {
1250
1251             /* REACTION-FIELD ELECTROSTATICS */
1252             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1253
1254             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1255
1256             fscal            = felec;
1257
1258             fscal            = _mm_and_pd(fscal,cutoff_mask);
1259
1260             /* Update vectorial force */
1261             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1262             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1263             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1264
1265             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1266             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1267             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1268
1269             }
1270
1271             /**************************
1272              * CALCULATE INTERACTIONS *
1273              **************************/
1274
1275             if (gmx_mm_any_lt(rsq21,rcutoff2))
1276             {
1277
1278             /* REACTION-FIELD ELECTROSTATICS */
1279             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1280
1281             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1282
1283             fscal            = felec;
1284
1285             fscal            = _mm_and_pd(fscal,cutoff_mask);
1286
1287             /* Update vectorial force */
1288             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1289             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1290             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1291
1292             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1293             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1294             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1295
1296             }
1297
1298             /**************************
1299              * CALCULATE INTERACTIONS *
1300              **************************/
1301
1302             if (gmx_mm_any_lt(rsq22,rcutoff2))
1303             {
1304
1305             /* REACTION-FIELD ELECTROSTATICS */
1306             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1307
1308             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1309
1310             fscal            = felec;
1311
1312             fscal            = _mm_and_pd(fscal,cutoff_mask);
1313
1314             /* Update vectorial force */
1315             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1316             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1317             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1318
1319             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1320             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1321             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1322
1323             }
1324
1325             /**************************
1326              * CALCULATE INTERACTIONS *
1327              **************************/
1328
1329             if (gmx_mm_any_lt(rsq23,rcutoff2))
1330             {
1331
1332             /* REACTION-FIELD ELECTROSTATICS */
1333             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1334
1335             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1336
1337             fscal            = felec;
1338
1339             fscal            = _mm_and_pd(fscal,cutoff_mask);
1340
1341             /* Update vectorial force */
1342             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1343             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1344             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1345
1346             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1347             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1348             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1349
1350             }
1351
1352             /**************************
1353              * CALCULATE INTERACTIONS *
1354              **************************/
1355
1356             if (gmx_mm_any_lt(rsq31,rcutoff2))
1357             {
1358
1359             /* REACTION-FIELD ELECTROSTATICS */
1360             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1361
1362             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1363
1364             fscal            = felec;
1365
1366             fscal            = _mm_and_pd(fscal,cutoff_mask);
1367
1368             /* Update vectorial force */
1369             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1370             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1371             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1372
1373             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1374             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1375             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1376
1377             }
1378
1379             /**************************
1380              * CALCULATE INTERACTIONS *
1381              **************************/
1382
1383             if (gmx_mm_any_lt(rsq32,rcutoff2))
1384             {
1385
1386             /* REACTION-FIELD ELECTROSTATICS */
1387             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1388
1389             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1390
1391             fscal            = felec;
1392
1393             fscal            = _mm_and_pd(fscal,cutoff_mask);
1394
1395             /* Update vectorial force */
1396             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1397             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1398             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1399
1400             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1401             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1402             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1403
1404             }
1405
1406             /**************************
1407              * CALCULATE INTERACTIONS *
1408              **************************/
1409
1410             if (gmx_mm_any_lt(rsq33,rcutoff2))
1411             {
1412
1413             /* REACTION-FIELD ELECTROSTATICS */
1414             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1415
1416             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1417
1418             fscal            = felec;
1419
1420             fscal            = _mm_and_pd(fscal,cutoff_mask);
1421
1422             /* Update vectorial force */
1423             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1424             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1425             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1426
1427             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1428             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1429             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1430
1431             }
1432
1433             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);
1434
1435             /* Inner loop uses 297 flops */
1436         }
1437
1438         if(jidx<j_index_end)
1439         {
1440
1441             jnrA             = jjnr[jidx];
1442             j_coord_offsetA  = DIM*jnrA;
1443
1444             /* load j atom coordinates */
1445             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
1446                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1447
1448             /* Calculate displacement vector */
1449             dx11             = _mm_sub_pd(ix1,jx1);
1450             dy11             = _mm_sub_pd(iy1,jy1);
1451             dz11             = _mm_sub_pd(iz1,jz1);
1452             dx12             = _mm_sub_pd(ix1,jx2);
1453             dy12             = _mm_sub_pd(iy1,jy2);
1454             dz12             = _mm_sub_pd(iz1,jz2);
1455             dx13             = _mm_sub_pd(ix1,jx3);
1456             dy13             = _mm_sub_pd(iy1,jy3);
1457             dz13             = _mm_sub_pd(iz1,jz3);
1458             dx21             = _mm_sub_pd(ix2,jx1);
1459             dy21             = _mm_sub_pd(iy2,jy1);
1460             dz21             = _mm_sub_pd(iz2,jz1);
1461             dx22             = _mm_sub_pd(ix2,jx2);
1462             dy22             = _mm_sub_pd(iy2,jy2);
1463             dz22             = _mm_sub_pd(iz2,jz2);
1464             dx23             = _mm_sub_pd(ix2,jx3);
1465             dy23             = _mm_sub_pd(iy2,jy3);
1466             dz23             = _mm_sub_pd(iz2,jz3);
1467             dx31             = _mm_sub_pd(ix3,jx1);
1468             dy31             = _mm_sub_pd(iy3,jy1);
1469             dz31             = _mm_sub_pd(iz3,jz1);
1470             dx32             = _mm_sub_pd(ix3,jx2);
1471             dy32             = _mm_sub_pd(iy3,jy2);
1472             dz32             = _mm_sub_pd(iz3,jz2);
1473             dx33             = _mm_sub_pd(ix3,jx3);
1474             dy33             = _mm_sub_pd(iy3,jy3);
1475             dz33             = _mm_sub_pd(iz3,jz3);
1476
1477             /* Calculate squared distance and things based on it */
1478             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1479             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1480             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1481             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1482             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1483             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1484             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1485             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1486             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1487
1488             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1489             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1490             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1491             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1492             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1493             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1494             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1495             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1496             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1497
1498             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1499             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1500             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1501             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1502             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1503             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1504             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1505             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1506             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1507
1508             fjx1             = _mm_setzero_pd();
1509             fjy1             = _mm_setzero_pd();
1510             fjz1             = _mm_setzero_pd();
1511             fjx2             = _mm_setzero_pd();
1512             fjy2             = _mm_setzero_pd();
1513             fjz2             = _mm_setzero_pd();
1514             fjx3             = _mm_setzero_pd();
1515             fjy3             = _mm_setzero_pd();
1516             fjz3             = _mm_setzero_pd();
1517
1518             /**************************
1519              * CALCULATE INTERACTIONS *
1520              **************************/
1521
1522             if (gmx_mm_any_lt(rsq11,rcutoff2))
1523             {
1524
1525             /* REACTION-FIELD ELECTROSTATICS */
1526             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1527
1528             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1529
1530             fscal            = felec;
1531
1532             fscal            = _mm_and_pd(fscal,cutoff_mask);
1533
1534             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1535
1536             /* Update vectorial force */
1537             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1538             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1539             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1540
1541             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1542             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1543             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1544
1545             }
1546
1547             /**************************
1548              * CALCULATE INTERACTIONS *
1549              **************************/
1550
1551             if (gmx_mm_any_lt(rsq12,rcutoff2))
1552             {
1553
1554             /* REACTION-FIELD ELECTROSTATICS */
1555             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1556
1557             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1558
1559             fscal            = felec;
1560
1561             fscal            = _mm_and_pd(fscal,cutoff_mask);
1562
1563             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1564
1565             /* Update vectorial force */
1566             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1567             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1568             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1569
1570             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1571             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1572             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1573
1574             }
1575
1576             /**************************
1577              * CALCULATE INTERACTIONS *
1578              **************************/
1579
1580             if (gmx_mm_any_lt(rsq13,rcutoff2))
1581             {
1582
1583             /* REACTION-FIELD ELECTROSTATICS */
1584             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1585
1586             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1587
1588             fscal            = felec;
1589
1590             fscal            = _mm_and_pd(fscal,cutoff_mask);
1591
1592             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1593
1594             /* Update vectorial force */
1595             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1596             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1597             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1598
1599             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1600             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1601             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1602
1603             }
1604
1605             /**************************
1606              * CALCULATE INTERACTIONS *
1607              **************************/
1608
1609             if (gmx_mm_any_lt(rsq21,rcutoff2))
1610             {
1611
1612             /* REACTION-FIELD ELECTROSTATICS */
1613             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1614
1615             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1616
1617             fscal            = felec;
1618
1619             fscal            = _mm_and_pd(fscal,cutoff_mask);
1620
1621             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1622
1623             /* Update vectorial force */
1624             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1625             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1626             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1627
1628             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1629             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1630             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1631
1632             }
1633
1634             /**************************
1635              * CALCULATE INTERACTIONS *
1636              **************************/
1637
1638             if (gmx_mm_any_lt(rsq22,rcutoff2))
1639             {
1640
1641             /* REACTION-FIELD ELECTROSTATICS */
1642             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1643
1644             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1645
1646             fscal            = felec;
1647
1648             fscal            = _mm_and_pd(fscal,cutoff_mask);
1649
1650             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1651
1652             /* Update vectorial force */
1653             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1654             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1655             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1656
1657             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1658             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1659             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1660
1661             }
1662
1663             /**************************
1664              * CALCULATE INTERACTIONS *
1665              **************************/
1666
1667             if (gmx_mm_any_lt(rsq23,rcutoff2))
1668             {
1669
1670             /* REACTION-FIELD ELECTROSTATICS */
1671             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1672
1673             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1674
1675             fscal            = felec;
1676
1677             fscal            = _mm_and_pd(fscal,cutoff_mask);
1678
1679             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1680
1681             /* Update vectorial force */
1682             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1683             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1684             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1685
1686             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1687             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1688             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1689
1690             }
1691
1692             /**************************
1693              * CALCULATE INTERACTIONS *
1694              **************************/
1695
1696             if (gmx_mm_any_lt(rsq31,rcutoff2))
1697             {
1698
1699             /* REACTION-FIELD ELECTROSTATICS */
1700             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1701
1702             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1703
1704             fscal            = felec;
1705
1706             fscal            = _mm_and_pd(fscal,cutoff_mask);
1707
1708             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1709
1710             /* Update vectorial force */
1711             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1712             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1713             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1714
1715             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1716             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1717             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1718
1719             }
1720
1721             /**************************
1722              * CALCULATE INTERACTIONS *
1723              **************************/
1724
1725             if (gmx_mm_any_lt(rsq32,rcutoff2))
1726             {
1727
1728             /* REACTION-FIELD ELECTROSTATICS */
1729             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1730
1731             cutoff_mask      = _mm_cmplt_pd(rsq32,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             /* Update vectorial force */
1740             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1741             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1742             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1743
1744             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1745             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1746             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1747
1748             }
1749
1750             /**************************
1751              * CALCULATE INTERACTIONS *
1752              **************************/
1753
1754             if (gmx_mm_any_lt(rsq33,rcutoff2))
1755             {
1756
1757             /* REACTION-FIELD ELECTROSTATICS */
1758             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1759
1760             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1761
1762             fscal            = felec;
1763
1764             fscal            = _mm_and_pd(fscal,cutoff_mask);
1765
1766             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1767
1768             /* Update vectorial force */
1769             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1770             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1771             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1772
1773             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1774             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1775             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1776
1777             }
1778
1779             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1780
1781             /* Inner loop uses 297 flops */
1782         }
1783
1784         /* End of innermost loop */
1785
1786         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1787                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1788
1789         /* Increment number of inner iterations */
1790         inneriter                  += j_index_end - j_index_start;
1791
1792         /* Outer loop uses 18 flops */
1793     }
1794
1795     /* Increment number of outer iterations */
1796     outeriter        += nri;
1797
1798     /* Update outer/inner flops */
1799
1800     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*297);
1801 }