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

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_double kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_double.h"
  48 #include "kernelutil_x86_avx_128_fma_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwNone_GeomW4W4_VF_avx_128_fma_double
  52  * Electrostatics interaction: ReactionField
  53  * VdW interaction:            None
  54  * Geometry:                   Water4-Water4
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecRFCut_VdwNone_GeomW4W4_VF_avx_128_fma_double
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB;
  75     int              j_coord_offsetA,j_coord_offsetB;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  80     int              vdwioffset1;
  81     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  82     int              vdwioffset2;
  83     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  84     int              vdwioffset3;
  85     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  86     int              vdwjidx1A,vdwjidx1B;
  87     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  88     int              vdwjidx2A,vdwjidx2B;
  89     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  90     int              vdwjidx3A,vdwjidx3B;
  91     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  92     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  93     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  94     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  95     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  96     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  97     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  98     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  99     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 100     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 101     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 102     real             *charge;
 103     __m128d          dummy_mask,cutoff_mask;
 104     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 105     __m128d          one     = _mm_set1_pd(1.0);
 106     __m128d          two     = _mm_set1_pd(2.0);
 107     x                = xx[0];
 108     f                = ff[0];
 109
 110     nri              = nlist->nri;
 111     iinr             = nlist->iinr;
 112     jindex           = nlist->jindex;
 113     jjnr             = nlist->jjnr;
 114     shiftidx         = nlist->shift;
 115     gid              = nlist->gid;
 116     shiftvec         = fr->shift_vec[0];
 117     fshift           = fr->fshift[0];
 118     facel            = _mm_set1_pd(fr->epsfac);
 119     charge           = mdatoms->chargeA;
 120     krf              = _mm_set1_pd(fr->ic->k_rf);
 121     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 122     crf              = _mm_set1_pd(fr->ic->c_rf);
 123
 124     /* Setup water-specific parameters */
 125     inr              = nlist->iinr[0];
 126     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 127     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 128     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 129
 130     jq1              = _mm_set1_pd(charge[inr+1]);
 131     jq2              = _mm_set1_pd(charge[inr+2]);
 132     jq3              = _mm_set1_pd(charge[inr+3]);
 133     qq11             = _mm_mul_pd(iq1,jq1);
 134     qq12             = _mm_mul_pd(iq1,jq2);
 135     qq13             = _mm_mul_pd(iq1,jq3);
 136     qq21             = _mm_mul_pd(iq2,jq1);
 137     qq22             = _mm_mul_pd(iq2,jq2);
 138     qq23             = _mm_mul_pd(iq2,jq3);
 139     qq31             = _mm_mul_pd(iq3,jq1);
 140     qq32             = _mm_mul_pd(iq3,jq2);
 141     qq33             = _mm_mul_pd(iq3,jq3);
 142
 143     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 144     rcutoff_scalar   = fr->rcoulomb;
 145     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 146     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 147
 148     /* Avoid stupid compiler warnings */
 149     jnrA = jnrB = 0;
 150     j_coord_offsetA = 0;
 151     j_coord_offsetB = 0;
 152
 153     outeriter        = 0;
 154     inneriter        = 0;
 155
 156     /* Start outer loop over neighborlists */
 157     for(iidx=0; iidx<nri; iidx++)
 158     {
 159         /* Load shift vector for this list */
 160         i_shift_offset   = DIM*shiftidx[iidx];
 161
 162         /* Load limits for loop over neighbors */
 163         j_index_start    = jindex[iidx];
 164         j_index_end      = jindex[iidx+1];
 165
 166         /* Get outer coordinate index */
 167         inr              = iinr[iidx];
 168         i_coord_offset   = DIM*inr;
 169
 170         /* Load i particle coords and add shift vector */
 171         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 172                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 173
 174         fix1             = _mm_setzero_pd();
 175         fiy1             = _mm_setzero_pd();
 176         fiz1             = _mm_setzero_pd();
 177         fix2             = _mm_setzero_pd();
 178         fiy2             = _mm_setzero_pd();
 179         fiz2             = _mm_setzero_pd();
 180         fix3             = _mm_setzero_pd();
 181         fiy3             = _mm_setzero_pd();
 182         fiz3             = _mm_setzero_pd();
 183
 184         /* Reset potential sums */
 185         velecsum         = _mm_setzero_pd();
 186
 187         /* Start inner kernel loop */
 188         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 189         {
 190
 191             /* Get j neighbor index, and coordinate index */
 192             jnrA             = jjnr[jidx];
 193             jnrB             = jjnr[jidx+1];
 194             j_coord_offsetA  = DIM*jnrA;
 195             j_coord_offsetB  = DIM*jnrB;
 196
 197             /* load j atom coordinates */
 198             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 199                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 200
 201             /* Calculate displacement vector */
 202             dx11             = _mm_sub_pd(ix1,jx1);
 203             dy11             = _mm_sub_pd(iy1,jy1);
 204             dz11             = _mm_sub_pd(iz1,jz1);
 205             dx12             = _mm_sub_pd(ix1,jx2);
 206             dy12             = _mm_sub_pd(iy1,jy2);
 207             dz12             = _mm_sub_pd(iz1,jz2);
 208             dx13             = _mm_sub_pd(ix1,jx3);
 209             dy13             = _mm_sub_pd(iy1,jy3);
 210             dz13             = _mm_sub_pd(iz1,jz3);
 211             dx21             = _mm_sub_pd(ix2,jx1);
 212             dy21             = _mm_sub_pd(iy2,jy1);
 213             dz21             = _mm_sub_pd(iz2,jz1);
 214             dx22             = _mm_sub_pd(ix2,jx2);
 215             dy22             = _mm_sub_pd(iy2,jy2);
 216             dz22             = _mm_sub_pd(iz2,jz2);
 217             dx23             = _mm_sub_pd(ix2,jx3);
 218             dy23             = _mm_sub_pd(iy2,jy3);
 219             dz23             = _mm_sub_pd(iz2,jz3);
 220             dx31             = _mm_sub_pd(ix3,jx1);
 221             dy31             = _mm_sub_pd(iy3,jy1);
 222             dz31             = _mm_sub_pd(iz3,jz1);
 223             dx32             = _mm_sub_pd(ix3,jx2);
 224             dy32             = _mm_sub_pd(iy3,jy2);
 225             dz32             = _mm_sub_pd(iz3,jz2);
 226             dx33             = _mm_sub_pd(ix3,jx3);
 227             dy33             = _mm_sub_pd(iy3,jy3);
 228             dz33             = _mm_sub_pd(iz3,jz3);
 229
 230             /* Calculate squared distance and things based on it */
 231             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 232             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 233             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 234             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 235             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 236             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 237             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 238             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 239             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 240
 241             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 242             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 243             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 244             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 245             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 246             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 247             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 248             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 249             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 250
 251             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 252             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 253             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 254             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 255             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 256             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 257             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 258             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 259             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 260
 261             fjx1             = _mm_setzero_pd();
 262             fjy1             = _mm_setzero_pd();
 263             fjz1             = _mm_setzero_pd();
 264             fjx2             = _mm_setzero_pd();
 265             fjy2             = _mm_setzero_pd();
 266             fjz2             = _mm_setzero_pd();
 267             fjx3             = _mm_setzero_pd();
 268             fjy3             = _mm_setzero_pd();
 269             fjz3             = _mm_setzero_pd();
 270
 271             /**************************
 272              * CALCULATE INTERACTIONS *
 273              **************************/
 274
 275             if (gmx_mm_any_lt(rsq11,rcutoff2))
 276             {
 277
 278             /* REACTION-FIELD ELECTROSTATICS */
 279             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 280             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 281
 282             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 283
 284             /* Update potential sum for this i atom from the interaction with this j atom. */
 285             velec            = _mm_and_pd(velec,cutoff_mask);
 286             velecsum         = _mm_add_pd(velecsum,velec);
 287
 288             fscal            = felec;
 289
 290             fscal            = _mm_and_pd(fscal,cutoff_mask);
 291
 292             /* Update vectorial force */
 293             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 294             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 295             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 296
 297             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 298             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 299             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 300
 301             }
 302
 303             /**************************
 304              * CALCULATE INTERACTIONS *
 305              **************************/
 306
 307             if (gmx_mm_any_lt(rsq12,rcutoff2))
 308             {
 309
 310             /* REACTION-FIELD ELECTROSTATICS */
 311             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 312             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 313
 314             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 315
 316             /* Update potential sum for this i atom from the interaction with this j atom. */
 317             velec            = _mm_and_pd(velec,cutoff_mask);
 318             velecsum         = _mm_add_pd(velecsum,velec);
 319
 320             fscal            = felec;
 321
 322             fscal            = _mm_and_pd(fscal,cutoff_mask);
 323
 324             /* Update vectorial force */
 325             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 326             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 327             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 328
 329             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 330             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 331             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 332
 333             }
 334
 335             /**************************
 336              * CALCULATE INTERACTIONS *
 337              **************************/
 338
 339             if (gmx_mm_any_lt(rsq13,rcutoff2))
 340             {
 341
 342             /* REACTION-FIELD ELECTROSTATICS */
 343             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 344             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 345
 346             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 347
 348             /* Update potential sum for this i atom from the interaction with this j atom. */
 349             velec            = _mm_and_pd(velec,cutoff_mask);
 350             velecsum         = _mm_add_pd(velecsum,velec);
 351
 352             fscal            = felec;
 353
 354             fscal            = _mm_and_pd(fscal,cutoff_mask);
 355
 356             /* Update vectorial force */
 357             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 358             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 359             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 360
 361             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 362             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 363             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 364
 365             }
 366
 367             /**************************
 368              * CALCULATE INTERACTIONS *
 369              **************************/
 370
 371             if (gmx_mm_any_lt(rsq21,rcutoff2))
 372             {
 373
 374             /* REACTION-FIELD ELECTROSTATICS */
 375             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 376             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 377
 378             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 379
 380             /* Update potential sum for this i atom from the interaction with this j atom. */
 381             velec            = _mm_and_pd(velec,cutoff_mask);
 382             velecsum         = _mm_add_pd(velecsum,velec);
 383
 384             fscal            = felec;
 385
 386             fscal            = _mm_and_pd(fscal,cutoff_mask);
 387
 388             /* Update vectorial force */
 389             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 390             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 391             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 392
 393             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 394             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 395             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 396
 397             }
 398
 399             /**************************
 400              * CALCULATE INTERACTIONS *
 401              **************************/
 402
 403             if (gmx_mm_any_lt(rsq22,rcutoff2))
 404             {
 405
 406             /* REACTION-FIELD ELECTROSTATICS */
 407             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 408             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 409
 410             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 411
 412             /* Update potential sum for this i atom from the interaction with this j atom. */
 413             velec            = _mm_and_pd(velec,cutoff_mask);
 414             velecsum         = _mm_add_pd(velecsum,velec);
 415
 416             fscal            = felec;
 417
 418             fscal            = _mm_and_pd(fscal,cutoff_mask);
 419
 420             /* Update vectorial force */
 421             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 422             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 423             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 424
 425             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 426             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 427             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 428
 429             }
 430
 431             /**************************
 432              * CALCULATE INTERACTIONS *
 433              **************************/
 434
 435             if (gmx_mm_any_lt(rsq23,rcutoff2))
 436             {
 437
 438             /* REACTION-FIELD ELECTROSTATICS */
 439             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 440             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 441
 442             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 443
 444             /* Update potential sum for this i atom from the interaction with this j atom. */
 445             velec            = _mm_and_pd(velec,cutoff_mask);
 446             velecsum         = _mm_add_pd(velecsum,velec);
 447
 448             fscal            = felec;
 449
 450             fscal            = _mm_and_pd(fscal,cutoff_mask);
 451
 452             /* Update vectorial force */
 453             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 454             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 455             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 456
 457             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 458             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 459             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 460
 461             }
 462
 463             /**************************
 464              * CALCULATE INTERACTIONS *
 465              **************************/
 466
 467             if (gmx_mm_any_lt(rsq31,rcutoff2))
 468             {
 469
 470             /* REACTION-FIELD ELECTROSTATICS */
 471             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
 472             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
 473
 474             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
 475
 476             /* Update potential sum for this i atom from the interaction with this j atom. */
 477             velec            = _mm_and_pd(velec,cutoff_mask);
 478             velecsum         = _mm_add_pd(velecsum,velec);
 479
 480             fscal            = felec;
 481
 482             fscal            = _mm_and_pd(fscal,cutoff_mask);
 483
 484             /* Update vectorial force */
 485             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 486             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 487             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 488
 489             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 490             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 491             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 492
 493             }
 494
 495             /**************************
 496              * CALCULATE INTERACTIONS *
 497              **************************/
 498
 499             if (gmx_mm_any_lt(rsq32,rcutoff2))
 500             {
 501
 502             /* REACTION-FIELD ELECTROSTATICS */
 503             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
 504             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
 505
 506             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
 507
 508             /* Update potential sum for this i atom from the interaction with this j atom. */
 509             velec            = _mm_and_pd(velec,cutoff_mask);
 510             velecsum         = _mm_add_pd(velecsum,velec);
 511
 512             fscal            = felec;
 513
 514             fscal            = _mm_and_pd(fscal,cutoff_mask);
 515
 516             /* Update vectorial force */
 517             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 518             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 519             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 520
 521             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 522             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 523             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 524
 525             }
 526
 527             /**************************
 528              * CALCULATE INTERACTIONS *
 529              **************************/
 530
 531             if (gmx_mm_any_lt(rsq33,rcutoff2))
 532             {
 533
 534             /* REACTION-FIELD ELECTROSTATICS */
 535             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
 536             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
 537
 538             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
 539
 540             /* Update potential sum for this i atom from the interaction with this j atom. */
 541             velec            = _mm_and_pd(velec,cutoff_mask);
 542             velecsum         = _mm_add_pd(velecsum,velec);
 543
 544             fscal            = felec;
 545
 546             fscal            = _mm_and_pd(fscal,cutoff_mask);
 547
 548             /* Update vectorial force */
 549             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 550             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 551             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 552
 553             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 554             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 555             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 556
 557             }
 558
 559             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);
 560
 561             /* Inner loop uses 351 flops */
 562         }
 563
 564         if(jidx<j_index_end)
 565         {
 566
 567             jnrA             = jjnr[jidx];
 568             j_coord_offsetA  = DIM*jnrA;
 569
 570             /* load j atom coordinates */
 571             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
 572                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 573
 574             /* Calculate displacement vector */
 575             dx11             = _mm_sub_pd(ix1,jx1);
 576             dy11             = _mm_sub_pd(iy1,jy1);
 577             dz11             = _mm_sub_pd(iz1,jz1);
 578             dx12             = _mm_sub_pd(ix1,jx2);
 579             dy12             = _mm_sub_pd(iy1,jy2);
 580             dz12             = _mm_sub_pd(iz1,jz2);
 581             dx13             = _mm_sub_pd(ix1,jx3);
 582             dy13             = _mm_sub_pd(iy1,jy3);
 583             dz13             = _mm_sub_pd(iz1,jz3);
 584             dx21             = _mm_sub_pd(ix2,jx1);
 585             dy21             = _mm_sub_pd(iy2,jy1);
 586             dz21             = _mm_sub_pd(iz2,jz1);
 587             dx22             = _mm_sub_pd(ix2,jx2);
 588             dy22             = _mm_sub_pd(iy2,jy2);
 589             dz22             = _mm_sub_pd(iz2,jz2);
 590             dx23             = _mm_sub_pd(ix2,jx3);
 591             dy23             = _mm_sub_pd(iy2,jy3);
 592             dz23             = _mm_sub_pd(iz2,jz3);
 593             dx31             = _mm_sub_pd(ix3,jx1);
 594             dy31             = _mm_sub_pd(iy3,jy1);
 595             dz31             = _mm_sub_pd(iz3,jz1);
 596             dx32             = _mm_sub_pd(ix3,jx2);
 597             dy32             = _mm_sub_pd(iy3,jy2);
 598             dz32             = _mm_sub_pd(iz3,jz2);
 599             dx33             = _mm_sub_pd(ix3,jx3);
 600             dy33             = _mm_sub_pd(iy3,jy3);
 601             dz33             = _mm_sub_pd(iz3,jz3);
 602
 603             /* Calculate squared distance and things based on it */
 604             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 605             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 606             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 607             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 608             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 609             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 610             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 611             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 612             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 613
 614             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 615             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 616             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 617             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 618             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 619             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 620             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 621             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 622             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 623
 624             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 625             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 626             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 627             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 628             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 629             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 630             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 631             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 632             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 633
 634             fjx1             = _mm_setzero_pd();
 635             fjy1             = _mm_setzero_pd();
 636             fjz1             = _mm_setzero_pd();
 637             fjx2             = _mm_setzero_pd();
 638             fjy2             = _mm_setzero_pd();
 639             fjz2             = _mm_setzero_pd();
 640             fjx3             = _mm_setzero_pd();
 641             fjy3             = _mm_setzero_pd();
 642             fjz3             = _mm_setzero_pd();
 643
 644             /**************************
 645              * CALCULATE INTERACTIONS *
 646              **************************/
 647
 648             if (gmx_mm_any_lt(rsq11,rcutoff2))
 649             {
 650
 651             /* REACTION-FIELD ELECTROSTATICS */
 652             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 653             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 654
 655             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 656
 657             /* Update potential sum for this i atom from the interaction with this j atom. */
 658             velec            = _mm_and_pd(velec,cutoff_mask);
 659             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 660             velecsum         = _mm_add_pd(velecsum,velec);
 661
 662             fscal            = felec;
 663
 664             fscal            = _mm_and_pd(fscal,cutoff_mask);
 665
 666             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 667
 668             /* Update vectorial force */
 669             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 670             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 671             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 672
 673             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 674             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 675             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 676
 677             }
 678
 679             /**************************
 680              * CALCULATE INTERACTIONS *
 681              **************************/
 682
 683             if (gmx_mm_any_lt(rsq12,rcutoff2))
 684             {
 685
 686             /* REACTION-FIELD ELECTROSTATICS */
 687             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 688             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 689
 690             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 691
 692             /* Update potential sum for this i atom from the interaction with this j atom. */
 693             velec            = _mm_and_pd(velec,cutoff_mask);
 694             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 695             velecsum         = _mm_add_pd(velecsum,velec);
 696
 697             fscal            = felec;
 698
 699             fscal            = _mm_and_pd(fscal,cutoff_mask);
 700
 701             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 702
 703             /* Update vectorial force */
 704             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 705             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 706             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 707
 708             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 709             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 710             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 711
 712             }
 713
 714             /**************************
 715              * CALCULATE INTERACTIONS *
 716              **************************/
 717
 718             if (gmx_mm_any_lt(rsq13,rcutoff2))
 719             {
 720
 721             /* REACTION-FIELD ELECTROSTATICS */
 722             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 723             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 724
 725             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 726
 727             /* Update potential sum for this i atom from the interaction with this j atom. */
 728             velec            = _mm_and_pd(velec,cutoff_mask);
 729             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 730             velecsum         = _mm_add_pd(velecsum,velec);
 731
 732             fscal            = felec;
 733
 734             fscal            = _mm_and_pd(fscal,cutoff_mask);
 735
 736             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 737
 738             /* Update vectorial force */
 739             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 740             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 741             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 742
 743             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 744             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 745             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 746
 747             }
 748
 749             /**************************
 750              * CALCULATE INTERACTIONS *
 751              **************************/
 752
 753             if (gmx_mm_any_lt(rsq21,rcutoff2))
 754             {
 755
 756             /* REACTION-FIELD ELECTROSTATICS */
 757             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 758             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 759
 760             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 761
 762             /* Update potential sum for this i atom from the interaction with this j atom. */
 763             velec            = _mm_and_pd(velec,cutoff_mask);
 764             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 765             velecsum         = _mm_add_pd(velecsum,velec);
 766
 767             fscal            = felec;
 768
 769             fscal            = _mm_and_pd(fscal,cutoff_mask);
 770
 771             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 772
 773             /* Update vectorial force */
 774             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 775             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 776             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 777
 778             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 779             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 780             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 781
 782             }
 783
 784             /**************************
 785              * CALCULATE INTERACTIONS *
 786              **************************/
 787
 788             if (gmx_mm_any_lt(rsq22,rcutoff2))
 789             {
 790
 791             /* REACTION-FIELD ELECTROSTATICS */
 792             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 793             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 794
 795             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 796
 797             /* Update potential sum for this i atom from the interaction with this j atom. */
 798             velec            = _mm_and_pd(velec,cutoff_mask);
 799             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 800             velecsum         = _mm_add_pd(velecsum,velec);
 801
 802             fscal            = felec;
 803
 804             fscal            = _mm_and_pd(fscal,cutoff_mask);
 805
 806             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 807
 808             /* Update vectorial force */
 809             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 810             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 811             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 812
 813             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 814             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 815             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 816
 817             }
 818
 819             /**************************
 820              * CALCULATE INTERACTIONS *
 821              **************************/
 822
 823             if (gmx_mm_any_lt(rsq23,rcutoff2))
 824             {
 825
 826             /* REACTION-FIELD ELECTROSTATICS */
 827             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 828             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 829
 830             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 831
 832             /* Update potential sum for this i atom from the interaction with this j atom. */
 833             velec            = _mm_and_pd(velec,cutoff_mask);
 834             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 835             velecsum         = _mm_add_pd(velecsum,velec);
 836
 837             fscal            = felec;
 838
 839             fscal            = _mm_and_pd(fscal,cutoff_mask);
 840
 841             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 842
 843             /* Update vectorial force */
 844             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 845             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 846             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 847
 848             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 849             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 850             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 851
 852             }
 853
 854             /**************************
 855              * CALCULATE INTERACTIONS *
 856              **************************/
 857
 858             if (gmx_mm_any_lt(rsq31,rcutoff2))
 859             {
 860
 861             /* REACTION-FIELD ELECTROSTATICS */
 862             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
 863             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
 864
 865             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
 866
 867             /* Update potential sum for this i atom from the interaction with this j atom. */
 868             velec            = _mm_and_pd(velec,cutoff_mask);
 869             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 870             velecsum         = _mm_add_pd(velecsum,velec);
 871
 872             fscal            = felec;
 873
 874             fscal            = _mm_and_pd(fscal,cutoff_mask);
 875
 876             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 877
 878             /* Update vectorial force */
 879             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 880             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 881             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 882
 883             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 884             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 885             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 886
 887             }
 888
 889             /**************************
 890              * CALCULATE INTERACTIONS *
 891              **************************/
 892
 893             if (gmx_mm_any_lt(rsq32,rcutoff2))
 894             {
 895
 896             /* REACTION-FIELD ELECTROSTATICS */
 897             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
 898             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
 899
 900             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
 901
 902             /* Update potential sum for this i atom from the interaction with this j atom. */
 903             velec            = _mm_and_pd(velec,cutoff_mask);
 904             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 905             velecsum         = _mm_add_pd(velecsum,velec);
 906
 907             fscal            = felec;
 908
 909             fscal            = _mm_and_pd(fscal,cutoff_mask);
 910
 911             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 912
 913             /* Update vectorial force */
 914             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 915             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 916             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 917
 918             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 919             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 920             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 921
 922             }
 923
 924             /**************************
 925              * CALCULATE INTERACTIONS *
 926              **************************/
 927
 928             if (gmx_mm_any_lt(rsq33,rcutoff2))
 929             {
 930
 931             /* REACTION-FIELD ELECTROSTATICS */
 932             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
 933             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
 934
 935             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
 936
 937             /* Update potential sum for this i atom from the interaction with this j atom. */
 938             velec            = _mm_and_pd(velec,cutoff_mask);
 939             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 940             velecsum         = _mm_add_pd(velecsum,velec);
 941
 942             fscal            = felec;
 943
 944             fscal            = _mm_and_pd(fscal,cutoff_mask);
 945
 946             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 947
 948             /* Update vectorial force */
 949             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 950             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 951             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 952
 953             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 954             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 955             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 956
 957             }
 958
 959             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 960
 961             /* Inner loop uses 351 flops */
 962         }
 963
 964         /* End of innermost loop */
 965
 966         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 967                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
 968
 969         ggid                        = gid[iidx];
 970         /* Update potential energies */
 971         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 972
 973         /* Increment number of inner iterations */
 974         inneriter                  += j_index_end - j_index_start;
 975
 976         /* Outer loop uses 19 flops */
 977     }
 978
 979     /* Increment number of outer iterations */
 980     outeriter        += nri;
 981
 982     /* Update outer/inner flops */
 983
 984     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*351);
 985 }
 986 /*
 987  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwNone_GeomW4W4_F_avx_128_fma_double
 988  * Electrostatics interaction: ReactionField
 989  * VdW interaction:            None
 990  * Geometry:                   Water4-Water4
 991  * Calculate force/pot:        Force
 992  */
 993 void
 994 nb_kernel_ElecRFCut_VdwNone_GeomW4W4_F_avx_128_fma_double
 995                     (t_nblist                    * gmx_restrict       nlist,
 996                      rvec                        * gmx_restrict          xx,
 997                      rvec                        * gmx_restrict          ff,
 998                      t_forcerec                  * gmx_restrict          fr,
 999                      t_mdatoms                   * gmx_restrict     mdatoms,
1000                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1001                      t_nrnb                      * gmx_restrict        nrnb)
1002 {
1003     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1004      * just 0 for non-waters.
1005      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1006      * jnr indices corresponding to data put in the four positions in the SIMD register.
1007      */
1008     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1009     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1010     int              jnrA,jnrB;
1011     int              j_coord_offsetA,j_coord_offsetB;
1012     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1013     real             rcutoff_scalar;
1014     real             *shiftvec,*fshift,*x,*f;
1015     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1016     int              vdwioffset1;
1017     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1018     int              vdwioffset2;
1019     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1020     int              vdwioffset3;
1021     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1022     int              vdwjidx1A,vdwjidx1B;
1023     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1024     int              vdwjidx2A,vdwjidx2B;
1025     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1026     int              vdwjidx3A,vdwjidx3B;
1027     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1028     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1029     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1030     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1031     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1032     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1033     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1034     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1035     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1036     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1037     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1038     real             *charge;
1039     __m128d          dummy_mask,cutoff_mask;
1040     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1041     __m128d          one     = _mm_set1_pd(1.0);
1042     __m128d          two     = _mm_set1_pd(2.0);
1043     x                = xx[0];
1044     f                = ff[0];
1045
1046     nri              = nlist->nri;
1047     iinr             = nlist->iinr;
1048     jindex           = nlist->jindex;
1049     jjnr             = nlist->jjnr;
1050     shiftidx         = nlist->shift;
1051     gid              = nlist->gid;
1052     shiftvec         = fr->shift_vec[0];
1053     fshift           = fr->fshift[0];
1054     facel            = _mm_set1_pd(fr->epsfac);
1055     charge           = mdatoms->chargeA;
1056     krf              = _mm_set1_pd(fr->ic->k_rf);
1057     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1058     crf              = _mm_set1_pd(fr->ic->c_rf);
1059
1060     /* Setup water-specific parameters */
1061     inr              = nlist->iinr[0];
1062     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1063     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1064     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1065
1066     jq1              = _mm_set1_pd(charge[inr+1]);
1067     jq2              = _mm_set1_pd(charge[inr+2]);
1068     jq3              = _mm_set1_pd(charge[inr+3]);
1069     qq11             = _mm_mul_pd(iq1,jq1);
1070     qq12             = _mm_mul_pd(iq1,jq2);
1071     qq13             = _mm_mul_pd(iq1,jq3);
1072     qq21             = _mm_mul_pd(iq2,jq1);
1073     qq22             = _mm_mul_pd(iq2,jq2);
1074     qq23             = _mm_mul_pd(iq2,jq3);
1075     qq31             = _mm_mul_pd(iq3,jq1);
1076     qq32             = _mm_mul_pd(iq3,jq2);
1077     qq33             = _mm_mul_pd(iq3,jq3);
1078
1079     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1080     rcutoff_scalar   = fr->rcoulomb;
1081     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1082     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1083
1084     /* Avoid stupid compiler warnings */
1085     jnrA = jnrB = 0;
1086     j_coord_offsetA = 0;
1087     j_coord_offsetB = 0;
1088
1089     outeriter        = 0;
1090     inneriter        = 0;
1091
1092     /* Start outer loop over neighborlists */
1093     for(iidx=0; iidx<nri; iidx++)
1094     {
1095         /* Load shift vector for this list */
1096         i_shift_offset   = DIM*shiftidx[iidx];
1097
1098         /* Load limits for loop over neighbors */
1099         j_index_start    = jindex[iidx];
1100         j_index_end      = jindex[iidx+1];
1101
1102         /* Get outer coordinate index */
1103         inr              = iinr[iidx];
1104         i_coord_offset   = DIM*inr;
1105
1106         /* Load i particle coords and add shift vector */
1107         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
1108                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1109
1110         fix1             = _mm_setzero_pd();
1111         fiy1             = _mm_setzero_pd();
1112         fiz1             = _mm_setzero_pd();
1113         fix2             = _mm_setzero_pd();
1114         fiy2             = _mm_setzero_pd();
1115         fiz2             = _mm_setzero_pd();
1116         fix3             = _mm_setzero_pd();
1117         fiy3             = _mm_setzero_pd();
1118         fiz3             = _mm_setzero_pd();
1119
1120         /* Start inner kernel loop */
1121         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1122         {
1123
1124             /* Get j neighbor index, and coordinate index */
1125             jnrA             = jjnr[jidx];
1126             jnrB             = jjnr[jidx+1];
1127             j_coord_offsetA  = DIM*jnrA;
1128             j_coord_offsetB  = DIM*jnrB;
1129
1130             /* load j atom coordinates */
1131             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1132                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1133
1134             /* Calculate displacement vector */
1135             dx11             = _mm_sub_pd(ix1,jx1);
1136             dy11             = _mm_sub_pd(iy1,jy1);
1137             dz11             = _mm_sub_pd(iz1,jz1);
1138             dx12             = _mm_sub_pd(ix1,jx2);
1139             dy12             = _mm_sub_pd(iy1,jy2);
1140             dz12             = _mm_sub_pd(iz1,jz2);
1141             dx13             = _mm_sub_pd(ix1,jx3);
1142             dy13             = _mm_sub_pd(iy1,jy3);
1143             dz13             = _mm_sub_pd(iz1,jz3);
1144             dx21             = _mm_sub_pd(ix2,jx1);
1145             dy21             = _mm_sub_pd(iy2,jy1);
1146             dz21             = _mm_sub_pd(iz2,jz1);
1147             dx22             = _mm_sub_pd(ix2,jx2);
1148             dy22             = _mm_sub_pd(iy2,jy2);
1149             dz22             = _mm_sub_pd(iz2,jz2);
1150             dx23             = _mm_sub_pd(ix2,jx3);
1151             dy23             = _mm_sub_pd(iy2,jy3);
1152             dz23             = _mm_sub_pd(iz2,jz3);
1153             dx31             = _mm_sub_pd(ix3,jx1);
1154             dy31             = _mm_sub_pd(iy3,jy1);
1155             dz31             = _mm_sub_pd(iz3,jz1);
1156             dx32             = _mm_sub_pd(ix3,jx2);
1157             dy32             = _mm_sub_pd(iy3,jy2);
1158             dz32             = _mm_sub_pd(iz3,jz2);
1159             dx33             = _mm_sub_pd(ix3,jx3);
1160             dy33             = _mm_sub_pd(iy3,jy3);
1161             dz33             = _mm_sub_pd(iz3,jz3);
1162
1163             /* Calculate squared distance and things based on it */
1164             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1165             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1166             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1167             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1168             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1169             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1170             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1171             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1172             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1173
1174             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1175             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1176             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1177             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1178             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1179             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1180             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1181             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1182             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1183
1184             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1185             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1186             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1187             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1188             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1189             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1190             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1191             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1192             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1193
1194             fjx1             = _mm_setzero_pd();
1195             fjy1             = _mm_setzero_pd();
1196             fjz1             = _mm_setzero_pd();
1197             fjx2             = _mm_setzero_pd();
1198             fjy2             = _mm_setzero_pd();
1199             fjz2             = _mm_setzero_pd();
1200             fjx3             = _mm_setzero_pd();
1201             fjy3             = _mm_setzero_pd();
1202             fjz3             = _mm_setzero_pd();
1203
1204             /**************************
1205              * CALCULATE INTERACTIONS *
1206              **************************/
1207
1208             if (gmx_mm_any_lt(rsq11,rcutoff2))
1209             {
1210
1211             /* REACTION-FIELD ELECTROSTATICS */
1212             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1213
1214             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1215
1216             fscal            = felec;
1217
1218             fscal            = _mm_and_pd(fscal,cutoff_mask);
1219
1220             /* Update vectorial force */
1221             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1222             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1223             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1224
1225             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1226             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1227             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1228
1229             }
1230
1231             /**************************
1232              * CALCULATE INTERACTIONS *
1233              **************************/
1234
1235             if (gmx_mm_any_lt(rsq12,rcutoff2))
1236             {
1237
1238             /* REACTION-FIELD ELECTROSTATICS */
1239             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1240
1241             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1242
1243             fscal            = felec;
1244
1245             fscal            = _mm_and_pd(fscal,cutoff_mask);
1246
1247             /* Update vectorial force */
1248             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1249             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1250             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1251
1252             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1253             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1254             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1255
1256             }
1257
1258             /**************************
1259              * CALCULATE INTERACTIONS *
1260              **************************/
1261
1262             if (gmx_mm_any_lt(rsq13,rcutoff2))
1263             {
1264
1265             /* REACTION-FIELD ELECTROSTATICS */
1266             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1267
1268             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1269
1270             fscal            = felec;
1271
1272             fscal            = _mm_and_pd(fscal,cutoff_mask);
1273
1274             /* Update vectorial force */
1275             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1276             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1277             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1278
1279             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1280             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1281             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1282
1283             }
1284
1285             /**************************
1286              * CALCULATE INTERACTIONS *
1287              **************************/
1288
1289             if (gmx_mm_any_lt(rsq21,rcutoff2))
1290             {
1291
1292             /* REACTION-FIELD ELECTROSTATICS */
1293             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1294
1295             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1296
1297             fscal            = felec;
1298
1299             fscal            = _mm_and_pd(fscal,cutoff_mask);
1300
1301             /* Update vectorial force */
1302             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1303             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1304             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1305
1306             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1307             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1308             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1309
1310             }
1311
1312             /**************************
1313              * CALCULATE INTERACTIONS *
1314              **************************/
1315
1316             if (gmx_mm_any_lt(rsq22,rcutoff2))
1317             {
1318
1319             /* REACTION-FIELD ELECTROSTATICS */
1320             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1321
1322             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1323
1324             fscal            = felec;
1325
1326             fscal            = _mm_and_pd(fscal,cutoff_mask);
1327
1328             /* Update vectorial force */
1329             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1330             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1331             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1332
1333             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1334             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1335             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1336
1337             }
1338
1339             /**************************
1340              * CALCULATE INTERACTIONS *
1341              **************************/
1342
1343             if (gmx_mm_any_lt(rsq23,rcutoff2))
1344             {
1345
1346             /* REACTION-FIELD ELECTROSTATICS */
1347             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1348
1349             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1350
1351             fscal            = felec;
1352
1353             fscal            = _mm_and_pd(fscal,cutoff_mask);
1354
1355             /* Update vectorial force */
1356             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1357             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1358             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1359
1360             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1361             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1362             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1363
1364             }
1365
1366             /**************************
1367              * CALCULATE INTERACTIONS *
1368              **************************/
1369
1370             if (gmx_mm_any_lt(rsq31,rcutoff2))
1371             {
1372
1373             /* REACTION-FIELD ELECTROSTATICS */
1374             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1375
1376             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1377
1378             fscal            = felec;
1379
1380             fscal            = _mm_and_pd(fscal,cutoff_mask);
1381
1382             /* Update vectorial force */
1383             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1384             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1385             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1386
1387             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1388             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1389             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1390
1391             }
1392
1393             /**************************
1394              * CALCULATE INTERACTIONS *
1395              **************************/
1396
1397             if (gmx_mm_any_lt(rsq32,rcutoff2))
1398             {
1399
1400             /* REACTION-FIELD ELECTROSTATICS */
1401             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1402
1403             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1404
1405             fscal            = felec;
1406
1407             fscal            = _mm_and_pd(fscal,cutoff_mask);
1408
1409             /* Update vectorial force */
1410             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1411             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1412             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1413
1414             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1415             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1416             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1417
1418             }
1419
1420             /**************************
1421              * CALCULATE INTERACTIONS *
1422              **************************/
1423
1424             if (gmx_mm_any_lt(rsq33,rcutoff2))
1425             {
1426
1427             /* REACTION-FIELD ELECTROSTATICS */
1428             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1429
1430             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1431
1432             fscal            = felec;
1433
1434             fscal            = _mm_and_pd(fscal,cutoff_mask);
1435
1436             /* Update vectorial force */
1437             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1438             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1439             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1440
1441             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1442             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1443             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1444
1445             }
1446
1447             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);
1448
1449             /* Inner loop uses 297 flops */
1450         }
1451
1452         if(jidx<j_index_end)
1453         {
1454
1455             jnrA             = jjnr[jidx];
1456             j_coord_offsetA  = DIM*jnrA;
1457
1458             /* load j atom coordinates */
1459             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
1460                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1461
1462             /* Calculate displacement vector */
1463             dx11             = _mm_sub_pd(ix1,jx1);
1464             dy11             = _mm_sub_pd(iy1,jy1);
1465             dz11             = _mm_sub_pd(iz1,jz1);
1466             dx12             = _mm_sub_pd(ix1,jx2);
1467             dy12             = _mm_sub_pd(iy1,jy2);
1468             dz12             = _mm_sub_pd(iz1,jz2);
1469             dx13             = _mm_sub_pd(ix1,jx3);
1470             dy13             = _mm_sub_pd(iy1,jy3);
1471             dz13             = _mm_sub_pd(iz1,jz3);
1472             dx21             = _mm_sub_pd(ix2,jx1);
1473             dy21             = _mm_sub_pd(iy2,jy1);
1474             dz21             = _mm_sub_pd(iz2,jz1);
1475             dx22             = _mm_sub_pd(ix2,jx2);
1476             dy22             = _mm_sub_pd(iy2,jy2);
1477             dz22             = _mm_sub_pd(iz2,jz2);
1478             dx23             = _mm_sub_pd(ix2,jx3);
1479             dy23             = _mm_sub_pd(iy2,jy3);
1480             dz23             = _mm_sub_pd(iz2,jz3);
1481             dx31             = _mm_sub_pd(ix3,jx1);
1482             dy31             = _mm_sub_pd(iy3,jy1);
1483             dz31             = _mm_sub_pd(iz3,jz1);
1484             dx32             = _mm_sub_pd(ix3,jx2);
1485             dy32             = _mm_sub_pd(iy3,jy2);
1486             dz32             = _mm_sub_pd(iz3,jz2);
1487             dx33             = _mm_sub_pd(ix3,jx3);
1488             dy33             = _mm_sub_pd(iy3,jy3);
1489             dz33             = _mm_sub_pd(iz3,jz3);
1490
1491             /* Calculate squared distance and things based on it */
1492             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1493             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1494             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1495             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1496             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1497             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1498             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1499             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1500             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1501
1502             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1503             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1504             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1505             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1506             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1507             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1508             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1509             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1510             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1511
1512             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1513             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1514             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1515             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1516             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1517             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1518             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1519             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1520             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1521
1522             fjx1             = _mm_setzero_pd();
1523             fjy1             = _mm_setzero_pd();
1524             fjz1             = _mm_setzero_pd();
1525             fjx2             = _mm_setzero_pd();
1526             fjy2             = _mm_setzero_pd();
1527             fjz2             = _mm_setzero_pd();
1528             fjx3             = _mm_setzero_pd();
1529             fjy3             = _mm_setzero_pd();
1530             fjz3             = _mm_setzero_pd();
1531
1532             /**************************
1533              * CALCULATE INTERACTIONS *
1534              **************************/
1535
1536             if (gmx_mm_any_lt(rsq11,rcutoff2))
1537             {
1538
1539             /* REACTION-FIELD ELECTROSTATICS */
1540             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1541
1542             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1543
1544             fscal            = felec;
1545
1546             fscal            = _mm_and_pd(fscal,cutoff_mask);
1547
1548             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1549
1550             /* Update vectorial force */
1551             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1552             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1553             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1554
1555             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1556             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1557             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1558
1559             }
1560
1561             /**************************
1562              * CALCULATE INTERACTIONS *
1563              **************************/
1564
1565             if (gmx_mm_any_lt(rsq12,rcutoff2))
1566             {
1567
1568             /* REACTION-FIELD ELECTROSTATICS */
1569             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1570
1571             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1572
1573             fscal            = felec;
1574
1575             fscal            = _mm_and_pd(fscal,cutoff_mask);
1576
1577             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1578
1579             /* Update vectorial force */
1580             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1581             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1582             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1583
1584             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1585             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1586             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1587
1588             }
1589
1590             /**************************
1591              * CALCULATE INTERACTIONS *
1592              **************************/
1593
1594             if (gmx_mm_any_lt(rsq13,rcutoff2))
1595             {
1596
1597             /* REACTION-FIELD ELECTROSTATICS */
1598             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1599
1600             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1601
1602             fscal            = felec;
1603
1604             fscal            = _mm_and_pd(fscal,cutoff_mask);
1605
1606             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1607
1608             /* Update vectorial force */
1609             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1610             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1611             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1612
1613             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1614             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1615             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1616
1617             }
1618
1619             /**************************
1620              * CALCULATE INTERACTIONS *
1621              **************************/
1622
1623             if (gmx_mm_any_lt(rsq21,rcutoff2))
1624             {
1625
1626             /* REACTION-FIELD ELECTROSTATICS */
1627             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1628
1629             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1630
1631             fscal            = felec;
1632
1633             fscal            = _mm_and_pd(fscal,cutoff_mask);
1634
1635             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1636
1637             /* Update vectorial force */
1638             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1639             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1640             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1641
1642             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1643             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1644             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1645
1646             }
1647
1648             /**************************
1649              * CALCULATE INTERACTIONS *
1650              **************************/
1651
1652             if (gmx_mm_any_lt(rsq22,rcutoff2))
1653             {
1654
1655             /* REACTION-FIELD ELECTROSTATICS */
1656             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1657
1658             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1659
1660             fscal            = felec;
1661
1662             fscal            = _mm_and_pd(fscal,cutoff_mask);
1663
1664             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1665
1666             /* Update vectorial force */
1667             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1668             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1669             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1670
1671             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1672             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1673             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1674
1675             }
1676
1677             /**************************
1678              * CALCULATE INTERACTIONS *
1679              **************************/
1680
1681             if (gmx_mm_any_lt(rsq23,rcutoff2))
1682             {
1683
1684             /* REACTION-FIELD ELECTROSTATICS */
1685             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1686
1687             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1688
1689             fscal            = felec;
1690
1691             fscal            = _mm_and_pd(fscal,cutoff_mask);
1692
1693             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1694
1695             /* Update vectorial force */
1696             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1697             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1698             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1699
1700             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1701             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1702             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1703
1704             }
1705
1706             /**************************
1707              * CALCULATE INTERACTIONS *
1708              **************************/
1709
1710             if (gmx_mm_any_lt(rsq31,rcutoff2))
1711             {
1712
1713             /* REACTION-FIELD ELECTROSTATICS */
1714             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1715
1716             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1717
1718             fscal            = felec;
1719
1720             fscal            = _mm_and_pd(fscal,cutoff_mask);
1721
1722             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1723
1724             /* Update vectorial force */
1725             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1726             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1727             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1728
1729             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1730             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1731             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1732
1733             }
1734
1735             /**************************
1736              * CALCULATE INTERACTIONS *
1737              **************************/
1738
1739             if (gmx_mm_any_lt(rsq32,rcutoff2))
1740             {
1741
1742             /* REACTION-FIELD ELECTROSTATICS */
1743             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1744
1745             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1746
1747             fscal            = felec;
1748
1749             fscal            = _mm_and_pd(fscal,cutoff_mask);
1750
1751             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1752
1753             /* Update vectorial force */
1754             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1755             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1756             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1757
1758             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1759             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1760             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1761
1762             }
1763
1764             /**************************
1765              * CALCULATE INTERACTIONS *
1766              **************************/
1767
1768             if (gmx_mm_any_lt(rsq33,rcutoff2))
1769             {
1770
1771             /* REACTION-FIELD ELECTROSTATICS */
1772             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1773
1774             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1775
1776             fscal            = felec;
1777
1778             fscal            = _mm_and_pd(fscal,cutoff_mask);
1779
1780             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1781
1782             /* Update vectorial force */
1783             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1784             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1785             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1786
1787             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1788             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1789             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1790
1791             }
1792
1793             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1794
1795             /* Inner loop uses 297 flops */
1796         }
1797
1798         /* End of innermost loop */
1799
1800         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1801                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1802
1803         /* Increment number of inner iterations */
1804         inneriter                  += j_index_end - j_index_start;
1805
1806         /* Outer loop uses 18 flops */
1807     }
1808
1809     /* Increment number of outer iterations */
1810     outeriter        += nri;
1811
1812     /* Update outer/inner flops */
1813
1814     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*297);
1815 }