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, 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 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_avx_128_fma_double.h"
  50 #include "kernelutil_x86_avx_128_fma_double.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwNone_GeomW4W4_VF_avx_128_fma_double
  54  * Electrostatics interaction: ReactionField
  55  * VdW interaction:            None
  56  * Geometry:                   Water4-Water4
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecRFCut_VdwNone_GeomW4W4_VF_avx_128_fma_double
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB;
  77     int              j_coord_offsetA,j_coord_offsetB;
  78     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  79     real             rcutoff_scalar;
  80     real             *shiftvec,*fshift,*x,*f;
  81     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset1;
  83     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  84     int              vdwioffset2;
  85     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  86     int              vdwioffset3;
  87     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  88     int              vdwjidx1A,vdwjidx1B;
  89     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  90     int              vdwjidx2A,vdwjidx2B;
  91     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  92     int              vdwjidx3A,vdwjidx3B;
  93     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  94     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  95     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  96     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  97     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  98     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  99     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 100     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 101     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 102     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 103     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 104     real             *charge;
 105     __m128d          dummy_mask,cutoff_mask;
 106     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 107     __m128d          one     = _mm_set1_pd(1.0);
 108     __m128d          two     = _mm_set1_pd(2.0);
 109     x                = xx[0];
 110     f                = ff[0];
 111
 112     nri              = nlist->nri;
 113     iinr             = nlist->iinr;
 114     jindex           = nlist->jindex;
 115     jjnr             = nlist->jjnr;
 116     shiftidx         = nlist->shift;
 117     gid              = nlist->gid;
 118     shiftvec         = fr->shift_vec[0];
 119     fshift           = fr->fshift[0];
 120     facel            = _mm_set1_pd(fr->epsfac);
 121     charge           = mdatoms->chargeA;
 122     krf              = _mm_set1_pd(fr->ic->k_rf);
 123     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 124     crf              = _mm_set1_pd(fr->ic->c_rf);
 125
 126     /* Setup water-specific parameters */
 127     inr              = nlist->iinr[0];
 128     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 129     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 130     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 131
 132     jq1              = _mm_set1_pd(charge[inr+1]);
 133     jq2              = _mm_set1_pd(charge[inr+2]);
 134     jq3              = _mm_set1_pd(charge[inr+3]);
 135     qq11             = _mm_mul_pd(iq1,jq1);
 136     qq12             = _mm_mul_pd(iq1,jq2);
 137     qq13             = _mm_mul_pd(iq1,jq3);
 138     qq21             = _mm_mul_pd(iq2,jq1);
 139     qq22             = _mm_mul_pd(iq2,jq2);
 140     qq23             = _mm_mul_pd(iq2,jq3);
 141     qq31             = _mm_mul_pd(iq3,jq1);
 142     qq32             = _mm_mul_pd(iq3,jq2);
 143     qq33             = _mm_mul_pd(iq3,jq3);
 144
 145     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 146     rcutoff_scalar   = fr->rcoulomb;
 147     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 148     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 149
 150     /* Avoid stupid compiler warnings */
 151     jnrA = jnrB = 0;
 152     j_coord_offsetA = 0;
 153     j_coord_offsetB = 0;
 154
 155     outeriter        = 0;
 156     inneriter        = 0;
 157
 158     /* Start outer loop over neighborlists */
 159     for(iidx=0; iidx<nri; iidx++)
 160     {
 161         /* Load shift vector for this list */
 162         i_shift_offset   = DIM*shiftidx[iidx];
 163
 164         /* Load limits for loop over neighbors */
 165         j_index_start    = jindex[iidx];
 166         j_index_end      = jindex[iidx+1];
 167
 168         /* Get outer coordinate index */
 169         inr              = iinr[iidx];
 170         i_coord_offset   = DIM*inr;
 171
 172         /* Load i particle coords and add shift vector */
 173         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 174                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 175
 176         fix1             = _mm_setzero_pd();
 177         fiy1             = _mm_setzero_pd();
 178         fiz1             = _mm_setzero_pd();
 179         fix2             = _mm_setzero_pd();
 180         fiy2             = _mm_setzero_pd();
 181         fiz2             = _mm_setzero_pd();
 182         fix3             = _mm_setzero_pd();
 183         fiy3             = _mm_setzero_pd();
 184         fiz3             = _mm_setzero_pd();
 185
 186         /* Reset potential sums */
 187         velecsum         = _mm_setzero_pd();
 188
 189         /* Start inner kernel loop */
 190         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 191         {
 192
 193             /* Get j neighbor index, and coordinate index */
 194             jnrA             = jjnr[jidx];
 195             jnrB             = jjnr[jidx+1];
 196             j_coord_offsetA  = DIM*jnrA;
 197             j_coord_offsetB  = DIM*jnrB;
 198
 199             /* load j atom coordinates */
 200             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 201                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 202
 203             /* Calculate displacement vector */
 204             dx11             = _mm_sub_pd(ix1,jx1);
 205             dy11             = _mm_sub_pd(iy1,jy1);
 206             dz11             = _mm_sub_pd(iz1,jz1);
 207             dx12             = _mm_sub_pd(ix1,jx2);
 208             dy12             = _mm_sub_pd(iy1,jy2);
 209             dz12             = _mm_sub_pd(iz1,jz2);
 210             dx13             = _mm_sub_pd(ix1,jx3);
 211             dy13             = _mm_sub_pd(iy1,jy3);
 212             dz13             = _mm_sub_pd(iz1,jz3);
 213             dx21             = _mm_sub_pd(ix2,jx1);
 214             dy21             = _mm_sub_pd(iy2,jy1);
 215             dz21             = _mm_sub_pd(iz2,jz1);
 216             dx22             = _mm_sub_pd(ix2,jx2);
 217             dy22             = _mm_sub_pd(iy2,jy2);
 218             dz22             = _mm_sub_pd(iz2,jz2);
 219             dx23             = _mm_sub_pd(ix2,jx3);
 220             dy23             = _mm_sub_pd(iy2,jy3);
 221             dz23             = _mm_sub_pd(iz2,jz3);
 222             dx31             = _mm_sub_pd(ix3,jx1);
 223             dy31             = _mm_sub_pd(iy3,jy1);
 224             dz31             = _mm_sub_pd(iz3,jz1);
 225             dx32             = _mm_sub_pd(ix3,jx2);
 226             dy32             = _mm_sub_pd(iy3,jy2);
 227             dz32             = _mm_sub_pd(iz3,jz2);
 228             dx33             = _mm_sub_pd(ix3,jx3);
 229             dy33             = _mm_sub_pd(iy3,jy3);
 230             dz33             = _mm_sub_pd(iz3,jz3);
 231
 232             /* Calculate squared distance and things based on it */
 233             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 234             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 235             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 236             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 237             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 238             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 239             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 240             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 241             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 242
 243             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 244             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 245             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 246             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 247             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 248             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 249             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 250             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 251             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 252
 253             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 254             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 255             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 256             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 257             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 258             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 259             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 260             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 261             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 262
 263             fjx1             = _mm_setzero_pd();
 264             fjy1             = _mm_setzero_pd();
 265             fjz1             = _mm_setzero_pd();
 266             fjx2             = _mm_setzero_pd();
 267             fjy2             = _mm_setzero_pd();
 268             fjz2             = _mm_setzero_pd();
 269             fjx3             = _mm_setzero_pd();
 270             fjy3             = _mm_setzero_pd();
 271             fjz3             = _mm_setzero_pd();
 272
 273             /**************************
 274              * CALCULATE INTERACTIONS *
 275              **************************/
 276
 277             if (gmx_mm_any_lt(rsq11,rcutoff2))
 278             {
 279
 280             /* REACTION-FIELD ELECTROSTATICS */
 281             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 282             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 283
 284             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 285
 286             /* Update potential sum for this i atom from the interaction with this j atom. */
 287             velec            = _mm_and_pd(velec,cutoff_mask);
 288             velecsum         = _mm_add_pd(velecsum,velec);
 289
 290             fscal            = felec;
 291
 292             fscal            = _mm_and_pd(fscal,cutoff_mask);
 293
 294             /* Update vectorial force */
 295             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 296             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 297             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 298
 299             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 300             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 301             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 302
 303             }
 304
 305             /**************************
 306              * CALCULATE INTERACTIONS *
 307              **************************/
 308
 309             if (gmx_mm_any_lt(rsq12,rcutoff2))
 310             {
 311
 312             /* REACTION-FIELD ELECTROSTATICS */
 313             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 314             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 315
 316             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 317
 318             /* Update potential sum for this i atom from the interaction with this j atom. */
 319             velec            = _mm_and_pd(velec,cutoff_mask);
 320             velecsum         = _mm_add_pd(velecsum,velec);
 321
 322             fscal            = felec;
 323
 324             fscal            = _mm_and_pd(fscal,cutoff_mask);
 325
 326             /* Update vectorial force */
 327             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 328             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 329             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 330
 331             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 332             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 333             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 334
 335             }
 336
 337             /**************************
 338              * CALCULATE INTERACTIONS *
 339              **************************/
 340
 341             if (gmx_mm_any_lt(rsq13,rcutoff2))
 342             {
 343
 344             /* REACTION-FIELD ELECTROSTATICS */
 345             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 346             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 347
 348             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 349
 350             /* Update potential sum for this i atom from the interaction with this j atom. */
 351             velec            = _mm_and_pd(velec,cutoff_mask);
 352             velecsum         = _mm_add_pd(velecsum,velec);
 353
 354             fscal            = felec;
 355
 356             fscal            = _mm_and_pd(fscal,cutoff_mask);
 357
 358             /* Update vectorial force */
 359             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 360             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 361             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 362
 363             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 364             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 365             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 366
 367             }
 368
 369             /**************************
 370              * CALCULATE INTERACTIONS *
 371              **************************/
 372
 373             if (gmx_mm_any_lt(rsq21,rcutoff2))
 374             {
 375
 376             /* REACTION-FIELD ELECTROSTATICS */
 377             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 378             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 379
 380             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 381
 382             /* Update potential sum for this i atom from the interaction with this j atom. */
 383             velec            = _mm_and_pd(velec,cutoff_mask);
 384             velecsum         = _mm_add_pd(velecsum,velec);
 385
 386             fscal            = felec;
 387
 388             fscal            = _mm_and_pd(fscal,cutoff_mask);
 389
 390             /* Update vectorial force */
 391             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 392             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 393             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 394
 395             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 396             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 397             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 398
 399             }
 400
 401             /**************************
 402              * CALCULATE INTERACTIONS *
 403              **************************/
 404
 405             if (gmx_mm_any_lt(rsq22,rcutoff2))
 406             {
 407
 408             /* REACTION-FIELD ELECTROSTATICS */
 409             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 410             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 411
 412             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 413
 414             /* Update potential sum for this i atom from the interaction with this j atom. */
 415             velec            = _mm_and_pd(velec,cutoff_mask);
 416             velecsum         = _mm_add_pd(velecsum,velec);
 417
 418             fscal            = felec;
 419
 420             fscal            = _mm_and_pd(fscal,cutoff_mask);
 421
 422             /* Update vectorial force */
 423             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 424             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 425             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 426
 427             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 428             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 429             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 430
 431             }
 432
 433             /**************************
 434              * CALCULATE INTERACTIONS *
 435              **************************/
 436
 437             if (gmx_mm_any_lt(rsq23,rcutoff2))
 438             {
 439
 440             /* REACTION-FIELD ELECTROSTATICS */
 441             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 442             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 443
 444             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 445
 446             /* Update potential sum for this i atom from the interaction with this j atom. */
 447             velec            = _mm_and_pd(velec,cutoff_mask);
 448             velecsum         = _mm_add_pd(velecsum,velec);
 449
 450             fscal            = felec;
 451
 452             fscal            = _mm_and_pd(fscal,cutoff_mask);
 453
 454             /* Update vectorial force */
 455             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 456             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 457             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 458
 459             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 460             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 461             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 462
 463             }
 464
 465             /**************************
 466              * CALCULATE INTERACTIONS *
 467              **************************/
 468
 469             if (gmx_mm_any_lt(rsq31,rcutoff2))
 470             {
 471
 472             /* REACTION-FIELD ELECTROSTATICS */
 473             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
 474             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
 475
 476             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
 477
 478             /* Update potential sum for this i atom from the interaction with this j atom. */
 479             velec            = _mm_and_pd(velec,cutoff_mask);
 480             velecsum         = _mm_add_pd(velecsum,velec);
 481
 482             fscal            = felec;
 483
 484             fscal            = _mm_and_pd(fscal,cutoff_mask);
 485
 486             /* Update vectorial force */
 487             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 488             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 489             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 490
 491             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 492             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 493             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 494
 495             }
 496
 497             /**************************
 498              * CALCULATE INTERACTIONS *
 499              **************************/
 500
 501             if (gmx_mm_any_lt(rsq32,rcutoff2))
 502             {
 503
 504             /* REACTION-FIELD ELECTROSTATICS */
 505             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
 506             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
 507
 508             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
 509
 510             /* Update potential sum for this i atom from the interaction with this j atom. */
 511             velec            = _mm_and_pd(velec,cutoff_mask);
 512             velecsum         = _mm_add_pd(velecsum,velec);
 513
 514             fscal            = felec;
 515
 516             fscal            = _mm_and_pd(fscal,cutoff_mask);
 517
 518             /* Update vectorial force */
 519             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 520             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 521             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 522
 523             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 524             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 525             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 526
 527             }
 528
 529             /**************************
 530              * CALCULATE INTERACTIONS *
 531              **************************/
 532
 533             if (gmx_mm_any_lt(rsq33,rcutoff2))
 534             {
 535
 536             /* REACTION-FIELD ELECTROSTATICS */
 537             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
 538             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
 539
 540             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
 541
 542             /* Update potential sum for this i atom from the interaction with this j atom. */
 543             velec            = _mm_and_pd(velec,cutoff_mask);
 544             velecsum         = _mm_add_pd(velecsum,velec);
 545
 546             fscal            = felec;
 547
 548             fscal            = _mm_and_pd(fscal,cutoff_mask);
 549
 550             /* Update vectorial force */
 551             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 552             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 553             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 554
 555             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 556             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 557             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 558
 559             }
 560
 561             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);
 562
 563             /* Inner loop uses 351 flops */
 564         }
 565
 566         if(jidx<j_index_end)
 567         {
 568
 569             jnrA             = jjnr[jidx];
 570             j_coord_offsetA  = DIM*jnrA;
 571
 572             /* load j atom coordinates */
 573             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
 574                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 575
 576             /* Calculate displacement vector */
 577             dx11             = _mm_sub_pd(ix1,jx1);
 578             dy11             = _mm_sub_pd(iy1,jy1);
 579             dz11             = _mm_sub_pd(iz1,jz1);
 580             dx12             = _mm_sub_pd(ix1,jx2);
 581             dy12             = _mm_sub_pd(iy1,jy2);
 582             dz12             = _mm_sub_pd(iz1,jz2);
 583             dx13             = _mm_sub_pd(ix1,jx3);
 584             dy13             = _mm_sub_pd(iy1,jy3);
 585             dz13             = _mm_sub_pd(iz1,jz3);
 586             dx21             = _mm_sub_pd(ix2,jx1);
 587             dy21             = _mm_sub_pd(iy2,jy1);
 588             dz21             = _mm_sub_pd(iz2,jz1);
 589             dx22             = _mm_sub_pd(ix2,jx2);
 590             dy22             = _mm_sub_pd(iy2,jy2);
 591             dz22             = _mm_sub_pd(iz2,jz2);
 592             dx23             = _mm_sub_pd(ix2,jx3);
 593             dy23             = _mm_sub_pd(iy2,jy3);
 594             dz23             = _mm_sub_pd(iz2,jz3);
 595             dx31             = _mm_sub_pd(ix3,jx1);
 596             dy31             = _mm_sub_pd(iy3,jy1);
 597             dz31             = _mm_sub_pd(iz3,jz1);
 598             dx32             = _mm_sub_pd(ix3,jx2);
 599             dy32             = _mm_sub_pd(iy3,jy2);
 600             dz32             = _mm_sub_pd(iz3,jz2);
 601             dx33             = _mm_sub_pd(ix3,jx3);
 602             dy33             = _mm_sub_pd(iy3,jy3);
 603             dz33             = _mm_sub_pd(iz3,jz3);
 604
 605             /* Calculate squared distance and things based on it */
 606             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 607             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 608             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 609             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 610             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 611             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 612             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 613             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 614             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 615
 616             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 617             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 618             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 619             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 620             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 621             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 622             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 623             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 624             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 625
 626             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 627             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 628             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 629             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 630             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 631             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 632             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 633             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 634             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 635
 636             fjx1             = _mm_setzero_pd();
 637             fjy1             = _mm_setzero_pd();
 638             fjz1             = _mm_setzero_pd();
 639             fjx2             = _mm_setzero_pd();
 640             fjy2             = _mm_setzero_pd();
 641             fjz2             = _mm_setzero_pd();
 642             fjx3             = _mm_setzero_pd();
 643             fjy3             = _mm_setzero_pd();
 644             fjz3             = _mm_setzero_pd();
 645
 646             /**************************
 647              * CALCULATE INTERACTIONS *
 648              **************************/
 649
 650             if (gmx_mm_any_lt(rsq11,rcutoff2))
 651             {
 652
 653             /* REACTION-FIELD ELECTROSTATICS */
 654             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 655             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 656
 657             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 658
 659             /* Update potential sum for this i atom from the interaction with this j atom. */
 660             velec            = _mm_and_pd(velec,cutoff_mask);
 661             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 662             velecsum         = _mm_add_pd(velecsum,velec);
 663
 664             fscal            = felec;
 665
 666             fscal            = _mm_and_pd(fscal,cutoff_mask);
 667
 668             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 669
 670             /* Update vectorial force */
 671             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 672             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 673             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 674
 675             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 676             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 677             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 678
 679             }
 680
 681             /**************************
 682              * CALCULATE INTERACTIONS *
 683              **************************/
 684
 685             if (gmx_mm_any_lt(rsq12,rcutoff2))
 686             {
 687
 688             /* REACTION-FIELD ELECTROSTATICS */
 689             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 690             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 691
 692             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 693
 694             /* Update potential sum for this i atom from the interaction with this j atom. */
 695             velec            = _mm_and_pd(velec,cutoff_mask);
 696             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 697             velecsum         = _mm_add_pd(velecsum,velec);
 698
 699             fscal            = felec;
 700
 701             fscal            = _mm_and_pd(fscal,cutoff_mask);
 702
 703             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 704
 705             /* Update vectorial force */
 706             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 707             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 708             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 709
 710             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 711             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 712             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 713
 714             }
 715
 716             /**************************
 717              * CALCULATE INTERACTIONS *
 718              **************************/
 719
 720             if (gmx_mm_any_lt(rsq13,rcutoff2))
 721             {
 722
 723             /* REACTION-FIELD ELECTROSTATICS */
 724             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 725             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 726
 727             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 728
 729             /* Update potential sum for this i atom from the interaction with this j atom. */
 730             velec            = _mm_and_pd(velec,cutoff_mask);
 731             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 732             velecsum         = _mm_add_pd(velecsum,velec);
 733
 734             fscal            = felec;
 735
 736             fscal            = _mm_and_pd(fscal,cutoff_mask);
 737
 738             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 739
 740             /* Update vectorial force */
 741             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 742             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 743             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 744
 745             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 746             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 747             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 748
 749             }
 750
 751             /**************************
 752              * CALCULATE INTERACTIONS *
 753              **************************/
 754
 755             if (gmx_mm_any_lt(rsq21,rcutoff2))
 756             {
 757
 758             /* REACTION-FIELD ELECTROSTATICS */
 759             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 760             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 761
 762             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 763
 764             /* Update potential sum for this i atom from the interaction with this j atom. */
 765             velec            = _mm_and_pd(velec,cutoff_mask);
 766             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 767             velecsum         = _mm_add_pd(velecsum,velec);
 768
 769             fscal            = felec;
 770
 771             fscal            = _mm_and_pd(fscal,cutoff_mask);
 772
 773             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 774
 775             /* Update vectorial force */
 776             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 777             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 778             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 779
 780             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 781             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 782             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 783
 784             }
 785
 786             /**************************
 787              * CALCULATE INTERACTIONS *
 788              **************************/
 789
 790             if (gmx_mm_any_lt(rsq22,rcutoff2))
 791             {
 792
 793             /* REACTION-FIELD ELECTROSTATICS */
 794             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 795             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 796
 797             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 798
 799             /* Update potential sum for this i atom from the interaction with this j atom. */
 800             velec            = _mm_and_pd(velec,cutoff_mask);
 801             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 802             velecsum         = _mm_add_pd(velecsum,velec);
 803
 804             fscal            = felec;
 805
 806             fscal            = _mm_and_pd(fscal,cutoff_mask);
 807
 808             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 809
 810             /* Update vectorial force */
 811             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 812             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 813             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 814
 815             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 816             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 817             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 818
 819             }
 820
 821             /**************************
 822              * CALCULATE INTERACTIONS *
 823              **************************/
 824
 825             if (gmx_mm_any_lt(rsq23,rcutoff2))
 826             {
 827
 828             /* REACTION-FIELD ELECTROSTATICS */
 829             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 830             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 831
 832             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 833
 834             /* Update potential sum for this i atom from the interaction with this j atom. */
 835             velec            = _mm_and_pd(velec,cutoff_mask);
 836             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 837             velecsum         = _mm_add_pd(velecsum,velec);
 838
 839             fscal            = felec;
 840
 841             fscal            = _mm_and_pd(fscal,cutoff_mask);
 842
 843             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 844
 845             /* Update vectorial force */
 846             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 847             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 848             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 849
 850             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 851             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 852             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 853
 854             }
 855
 856             /**************************
 857              * CALCULATE INTERACTIONS *
 858              **************************/
 859
 860             if (gmx_mm_any_lt(rsq31,rcutoff2))
 861             {
 862
 863             /* REACTION-FIELD ELECTROSTATICS */
 864             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
 865             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
 866
 867             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
 868
 869             /* Update potential sum for this i atom from the interaction with this j atom. */
 870             velec            = _mm_and_pd(velec,cutoff_mask);
 871             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 872             velecsum         = _mm_add_pd(velecsum,velec);
 873
 874             fscal            = felec;
 875
 876             fscal            = _mm_and_pd(fscal,cutoff_mask);
 877
 878             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 879
 880             /* Update vectorial force */
 881             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 882             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 883             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 884
 885             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 886             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 887             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 888
 889             }
 890
 891             /**************************
 892              * CALCULATE INTERACTIONS *
 893              **************************/
 894
 895             if (gmx_mm_any_lt(rsq32,rcutoff2))
 896             {
 897
 898             /* REACTION-FIELD ELECTROSTATICS */
 899             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
 900             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
 901
 902             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
 903
 904             /* Update potential sum for this i atom from the interaction with this j atom. */
 905             velec            = _mm_and_pd(velec,cutoff_mask);
 906             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 907             velecsum         = _mm_add_pd(velecsum,velec);
 908
 909             fscal            = felec;
 910
 911             fscal            = _mm_and_pd(fscal,cutoff_mask);
 912
 913             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 914
 915             /* Update vectorial force */
 916             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 917             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 918             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 919
 920             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 921             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 922             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 923
 924             }
 925
 926             /**************************
 927              * CALCULATE INTERACTIONS *
 928              **************************/
 929
 930             if (gmx_mm_any_lt(rsq33,rcutoff2))
 931             {
 932
 933             /* REACTION-FIELD ELECTROSTATICS */
 934             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
 935             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
 936
 937             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
 938
 939             /* Update potential sum for this i atom from the interaction with this j atom. */
 940             velec            = _mm_and_pd(velec,cutoff_mask);
 941             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 942             velecsum         = _mm_add_pd(velecsum,velec);
 943
 944             fscal            = felec;
 945
 946             fscal            = _mm_and_pd(fscal,cutoff_mask);
 947
 948             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 949
 950             /* Update vectorial force */
 951             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 952             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 953             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 954
 955             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 956             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 957             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 958
 959             }
 960
 961             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 962
 963             /* Inner loop uses 351 flops */
 964         }
 965
 966         /* End of innermost loop */
 967
 968         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 969                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
 970
 971         ggid                        = gid[iidx];
 972         /* Update potential energies */
 973         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 974
 975         /* Increment number of inner iterations */
 976         inneriter                  += j_index_end - j_index_start;
 977
 978         /* Outer loop uses 19 flops */
 979     }
 980
 981     /* Increment number of outer iterations */
 982     outeriter        += nri;
 983
 984     /* Update outer/inner flops */
 985
 986     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*351);
 987 }
 988 /*
 989  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwNone_GeomW4W4_F_avx_128_fma_double
 990  * Electrostatics interaction: ReactionField
 991  * VdW interaction:            None
 992  * Geometry:                   Water4-Water4
 993  * Calculate force/pot:        Force
 994  */
 995 void
 996 nb_kernel_ElecRFCut_VdwNone_GeomW4W4_F_avx_128_fma_double
 997                     (t_nblist                    * gmx_restrict       nlist,
 998                      rvec                        * gmx_restrict          xx,
 999                      rvec                        * gmx_restrict          ff,
1000                      t_forcerec                  * gmx_restrict          fr,
1001                      t_mdatoms                   * gmx_restrict     mdatoms,
1002                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1003                      t_nrnb                      * gmx_restrict        nrnb)
1004 {
1005     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1006      * just 0 for non-waters.
1007      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1008      * jnr indices corresponding to data put in the four positions in the SIMD register.
1009      */
1010     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1011     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1012     int              jnrA,jnrB;
1013     int              j_coord_offsetA,j_coord_offsetB;
1014     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1015     real             rcutoff_scalar;
1016     real             *shiftvec,*fshift,*x,*f;
1017     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1018     int              vdwioffset1;
1019     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1020     int              vdwioffset2;
1021     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1022     int              vdwioffset3;
1023     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1024     int              vdwjidx1A,vdwjidx1B;
1025     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1026     int              vdwjidx2A,vdwjidx2B;
1027     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1028     int              vdwjidx3A,vdwjidx3B;
1029     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1030     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1031     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1032     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1033     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1034     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1035     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1036     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1037     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1038     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1039     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1040     real             *charge;
1041     __m128d          dummy_mask,cutoff_mask;
1042     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1043     __m128d          one     = _mm_set1_pd(1.0);
1044     __m128d          two     = _mm_set1_pd(2.0);
1045     x                = xx[0];
1046     f                = ff[0];
1047
1048     nri              = nlist->nri;
1049     iinr             = nlist->iinr;
1050     jindex           = nlist->jindex;
1051     jjnr             = nlist->jjnr;
1052     shiftidx         = nlist->shift;
1053     gid              = nlist->gid;
1054     shiftvec         = fr->shift_vec[0];
1055     fshift           = fr->fshift[0];
1056     facel            = _mm_set1_pd(fr->epsfac);
1057     charge           = mdatoms->chargeA;
1058     krf              = _mm_set1_pd(fr->ic->k_rf);
1059     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1060     crf              = _mm_set1_pd(fr->ic->c_rf);
1061
1062     /* Setup water-specific parameters */
1063     inr              = nlist->iinr[0];
1064     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1065     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1066     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1067
1068     jq1              = _mm_set1_pd(charge[inr+1]);
1069     jq2              = _mm_set1_pd(charge[inr+2]);
1070     jq3              = _mm_set1_pd(charge[inr+3]);
1071     qq11             = _mm_mul_pd(iq1,jq1);
1072     qq12             = _mm_mul_pd(iq1,jq2);
1073     qq13             = _mm_mul_pd(iq1,jq3);
1074     qq21             = _mm_mul_pd(iq2,jq1);
1075     qq22             = _mm_mul_pd(iq2,jq2);
1076     qq23             = _mm_mul_pd(iq2,jq3);
1077     qq31             = _mm_mul_pd(iq3,jq1);
1078     qq32             = _mm_mul_pd(iq3,jq2);
1079     qq33             = _mm_mul_pd(iq3,jq3);
1080
1081     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1082     rcutoff_scalar   = fr->rcoulomb;
1083     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1084     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1085
1086     /* Avoid stupid compiler warnings */
1087     jnrA = jnrB = 0;
1088     j_coord_offsetA = 0;
1089     j_coord_offsetB = 0;
1090
1091     outeriter        = 0;
1092     inneriter        = 0;
1093
1094     /* Start outer loop over neighborlists */
1095     for(iidx=0; iidx<nri; iidx++)
1096     {
1097         /* Load shift vector for this list */
1098         i_shift_offset   = DIM*shiftidx[iidx];
1099
1100         /* Load limits for loop over neighbors */
1101         j_index_start    = jindex[iidx];
1102         j_index_end      = jindex[iidx+1];
1103
1104         /* Get outer coordinate index */
1105         inr              = iinr[iidx];
1106         i_coord_offset   = DIM*inr;
1107
1108         /* Load i particle coords and add shift vector */
1109         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
1110                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1111
1112         fix1             = _mm_setzero_pd();
1113         fiy1             = _mm_setzero_pd();
1114         fiz1             = _mm_setzero_pd();
1115         fix2             = _mm_setzero_pd();
1116         fiy2             = _mm_setzero_pd();
1117         fiz2             = _mm_setzero_pd();
1118         fix3             = _mm_setzero_pd();
1119         fiy3             = _mm_setzero_pd();
1120         fiz3             = _mm_setzero_pd();
1121
1122         /* Start inner kernel loop */
1123         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1124         {
1125
1126             /* Get j neighbor index, and coordinate index */
1127             jnrA             = jjnr[jidx];
1128             jnrB             = jjnr[jidx+1];
1129             j_coord_offsetA  = DIM*jnrA;
1130             j_coord_offsetB  = DIM*jnrB;
1131
1132             /* load j atom coordinates */
1133             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1134                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1135
1136             /* Calculate displacement vector */
1137             dx11             = _mm_sub_pd(ix1,jx1);
1138             dy11             = _mm_sub_pd(iy1,jy1);
1139             dz11             = _mm_sub_pd(iz1,jz1);
1140             dx12             = _mm_sub_pd(ix1,jx2);
1141             dy12             = _mm_sub_pd(iy1,jy2);
1142             dz12             = _mm_sub_pd(iz1,jz2);
1143             dx13             = _mm_sub_pd(ix1,jx3);
1144             dy13             = _mm_sub_pd(iy1,jy3);
1145             dz13             = _mm_sub_pd(iz1,jz3);
1146             dx21             = _mm_sub_pd(ix2,jx1);
1147             dy21             = _mm_sub_pd(iy2,jy1);
1148             dz21             = _mm_sub_pd(iz2,jz1);
1149             dx22             = _mm_sub_pd(ix2,jx2);
1150             dy22             = _mm_sub_pd(iy2,jy2);
1151             dz22             = _mm_sub_pd(iz2,jz2);
1152             dx23             = _mm_sub_pd(ix2,jx3);
1153             dy23             = _mm_sub_pd(iy2,jy3);
1154             dz23             = _mm_sub_pd(iz2,jz3);
1155             dx31             = _mm_sub_pd(ix3,jx1);
1156             dy31             = _mm_sub_pd(iy3,jy1);
1157             dz31             = _mm_sub_pd(iz3,jz1);
1158             dx32             = _mm_sub_pd(ix3,jx2);
1159             dy32             = _mm_sub_pd(iy3,jy2);
1160             dz32             = _mm_sub_pd(iz3,jz2);
1161             dx33             = _mm_sub_pd(ix3,jx3);
1162             dy33             = _mm_sub_pd(iy3,jy3);
1163             dz33             = _mm_sub_pd(iz3,jz3);
1164
1165             /* Calculate squared distance and things based on it */
1166             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1167             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1168             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1169             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1170             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1171             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1172             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1173             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1174             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1175
1176             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1177             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1178             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1179             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1180             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1181             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1182             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1183             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1184             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1185
1186             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1187             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1188             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1189             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1190             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1191             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1192             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1193             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1194             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1195
1196             fjx1             = _mm_setzero_pd();
1197             fjy1             = _mm_setzero_pd();
1198             fjz1             = _mm_setzero_pd();
1199             fjx2             = _mm_setzero_pd();
1200             fjy2             = _mm_setzero_pd();
1201             fjz2             = _mm_setzero_pd();
1202             fjx3             = _mm_setzero_pd();
1203             fjy3             = _mm_setzero_pd();
1204             fjz3             = _mm_setzero_pd();
1205
1206             /**************************
1207              * CALCULATE INTERACTIONS *
1208              **************************/
1209
1210             if (gmx_mm_any_lt(rsq11,rcutoff2))
1211             {
1212
1213             /* REACTION-FIELD ELECTROSTATICS */
1214             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1215
1216             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1217
1218             fscal            = felec;
1219
1220             fscal            = _mm_and_pd(fscal,cutoff_mask);
1221
1222             /* Update vectorial force */
1223             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1224             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1225             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1226
1227             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1228             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1229             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1230
1231             }
1232
1233             /**************************
1234              * CALCULATE INTERACTIONS *
1235              **************************/
1236
1237             if (gmx_mm_any_lt(rsq12,rcutoff2))
1238             {
1239
1240             /* REACTION-FIELD ELECTROSTATICS */
1241             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1242
1243             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1244
1245             fscal            = felec;
1246
1247             fscal            = _mm_and_pd(fscal,cutoff_mask);
1248
1249             /* Update vectorial force */
1250             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1251             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1252             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1253
1254             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1255             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1256             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1257
1258             }
1259
1260             /**************************
1261              * CALCULATE INTERACTIONS *
1262              **************************/
1263
1264             if (gmx_mm_any_lt(rsq13,rcutoff2))
1265             {
1266
1267             /* REACTION-FIELD ELECTROSTATICS */
1268             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1269
1270             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1271
1272             fscal            = felec;
1273
1274             fscal            = _mm_and_pd(fscal,cutoff_mask);
1275
1276             /* Update vectorial force */
1277             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1278             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1279             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1280
1281             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1282             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1283             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1284
1285             }
1286
1287             /**************************
1288              * CALCULATE INTERACTIONS *
1289              **************************/
1290
1291             if (gmx_mm_any_lt(rsq21,rcutoff2))
1292             {
1293
1294             /* REACTION-FIELD ELECTROSTATICS */
1295             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1296
1297             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1298
1299             fscal            = felec;
1300
1301             fscal            = _mm_and_pd(fscal,cutoff_mask);
1302
1303             /* Update vectorial force */
1304             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1305             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1306             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1307
1308             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1309             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1310             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1311
1312             }
1313
1314             /**************************
1315              * CALCULATE INTERACTIONS *
1316              **************************/
1317
1318             if (gmx_mm_any_lt(rsq22,rcutoff2))
1319             {
1320
1321             /* REACTION-FIELD ELECTROSTATICS */
1322             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1323
1324             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1325
1326             fscal            = felec;
1327
1328             fscal            = _mm_and_pd(fscal,cutoff_mask);
1329
1330             /* Update vectorial force */
1331             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1332             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1333             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1334
1335             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1336             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1337             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1338
1339             }
1340
1341             /**************************
1342              * CALCULATE INTERACTIONS *
1343              **************************/
1344
1345             if (gmx_mm_any_lt(rsq23,rcutoff2))
1346             {
1347
1348             /* REACTION-FIELD ELECTROSTATICS */
1349             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1350
1351             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1352
1353             fscal            = felec;
1354
1355             fscal            = _mm_and_pd(fscal,cutoff_mask);
1356
1357             /* Update vectorial force */
1358             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1359             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1360             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1361
1362             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1363             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1364             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1365
1366             }
1367
1368             /**************************
1369              * CALCULATE INTERACTIONS *
1370              **************************/
1371
1372             if (gmx_mm_any_lt(rsq31,rcutoff2))
1373             {
1374
1375             /* REACTION-FIELD ELECTROSTATICS */
1376             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1377
1378             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1379
1380             fscal            = felec;
1381
1382             fscal            = _mm_and_pd(fscal,cutoff_mask);
1383
1384             /* Update vectorial force */
1385             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1386             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1387             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1388
1389             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1390             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1391             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1392
1393             }
1394
1395             /**************************
1396              * CALCULATE INTERACTIONS *
1397              **************************/
1398
1399             if (gmx_mm_any_lt(rsq32,rcutoff2))
1400             {
1401
1402             /* REACTION-FIELD ELECTROSTATICS */
1403             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1404
1405             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1406
1407             fscal            = felec;
1408
1409             fscal            = _mm_and_pd(fscal,cutoff_mask);
1410
1411             /* Update vectorial force */
1412             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1413             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1414             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1415
1416             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1417             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1418             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1419
1420             }
1421
1422             /**************************
1423              * CALCULATE INTERACTIONS *
1424              **************************/
1425
1426             if (gmx_mm_any_lt(rsq33,rcutoff2))
1427             {
1428
1429             /* REACTION-FIELD ELECTROSTATICS */
1430             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1431
1432             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1433
1434             fscal            = felec;
1435
1436             fscal            = _mm_and_pd(fscal,cutoff_mask);
1437
1438             /* Update vectorial force */
1439             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1440             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1441             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1442
1443             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1444             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1445             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1446
1447             }
1448
1449             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);
1450
1451             /* Inner loop uses 297 flops */
1452         }
1453
1454         if(jidx<j_index_end)
1455         {
1456
1457             jnrA             = jjnr[jidx];
1458             j_coord_offsetA  = DIM*jnrA;
1459
1460             /* load j atom coordinates */
1461             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
1462                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1463
1464             /* Calculate displacement vector */
1465             dx11             = _mm_sub_pd(ix1,jx1);
1466             dy11             = _mm_sub_pd(iy1,jy1);
1467             dz11             = _mm_sub_pd(iz1,jz1);
1468             dx12             = _mm_sub_pd(ix1,jx2);
1469             dy12             = _mm_sub_pd(iy1,jy2);
1470             dz12             = _mm_sub_pd(iz1,jz2);
1471             dx13             = _mm_sub_pd(ix1,jx3);
1472             dy13             = _mm_sub_pd(iy1,jy3);
1473             dz13             = _mm_sub_pd(iz1,jz3);
1474             dx21             = _mm_sub_pd(ix2,jx1);
1475             dy21             = _mm_sub_pd(iy2,jy1);
1476             dz21             = _mm_sub_pd(iz2,jz1);
1477             dx22             = _mm_sub_pd(ix2,jx2);
1478             dy22             = _mm_sub_pd(iy2,jy2);
1479             dz22             = _mm_sub_pd(iz2,jz2);
1480             dx23             = _mm_sub_pd(ix2,jx3);
1481             dy23             = _mm_sub_pd(iy2,jy3);
1482             dz23             = _mm_sub_pd(iz2,jz3);
1483             dx31             = _mm_sub_pd(ix3,jx1);
1484             dy31             = _mm_sub_pd(iy3,jy1);
1485             dz31             = _mm_sub_pd(iz3,jz1);
1486             dx32             = _mm_sub_pd(ix3,jx2);
1487             dy32             = _mm_sub_pd(iy3,jy2);
1488             dz32             = _mm_sub_pd(iz3,jz2);
1489             dx33             = _mm_sub_pd(ix3,jx3);
1490             dy33             = _mm_sub_pd(iy3,jy3);
1491             dz33             = _mm_sub_pd(iz3,jz3);
1492
1493             /* Calculate squared distance and things based on it */
1494             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1495             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1496             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1497             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1498             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1499             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1500             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1501             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1502             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1503
1504             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1505             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1506             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1507             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1508             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1509             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1510             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1511             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1512             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1513
1514             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1515             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1516             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1517             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1518             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1519             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1520             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1521             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1522             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1523
1524             fjx1             = _mm_setzero_pd();
1525             fjy1             = _mm_setzero_pd();
1526             fjz1             = _mm_setzero_pd();
1527             fjx2             = _mm_setzero_pd();
1528             fjy2             = _mm_setzero_pd();
1529             fjz2             = _mm_setzero_pd();
1530             fjx3             = _mm_setzero_pd();
1531             fjy3             = _mm_setzero_pd();
1532             fjz3             = _mm_setzero_pd();
1533
1534             /**************************
1535              * CALCULATE INTERACTIONS *
1536              **************************/
1537
1538             if (gmx_mm_any_lt(rsq11,rcutoff2))
1539             {
1540
1541             /* REACTION-FIELD ELECTROSTATICS */
1542             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1543
1544             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1545
1546             fscal            = felec;
1547
1548             fscal            = _mm_and_pd(fscal,cutoff_mask);
1549
1550             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1551
1552             /* Update vectorial force */
1553             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1554             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1555             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1556
1557             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1558             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1559             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1560
1561             }
1562
1563             /**************************
1564              * CALCULATE INTERACTIONS *
1565              **************************/
1566
1567             if (gmx_mm_any_lt(rsq12,rcutoff2))
1568             {
1569
1570             /* REACTION-FIELD ELECTROSTATICS */
1571             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1572
1573             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1574
1575             fscal            = felec;
1576
1577             fscal            = _mm_and_pd(fscal,cutoff_mask);
1578
1579             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1580
1581             /* Update vectorial force */
1582             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1583             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1584             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1585
1586             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1587             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1588             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1589
1590             }
1591
1592             /**************************
1593              * CALCULATE INTERACTIONS *
1594              **************************/
1595
1596             if (gmx_mm_any_lt(rsq13,rcutoff2))
1597             {
1598
1599             /* REACTION-FIELD ELECTROSTATICS */
1600             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1601
1602             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1603
1604             fscal            = felec;
1605
1606             fscal            = _mm_and_pd(fscal,cutoff_mask);
1607
1608             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1609
1610             /* Update vectorial force */
1611             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1612             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1613             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1614
1615             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1616             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1617             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1618
1619             }
1620
1621             /**************************
1622              * CALCULATE INTERACTIONS *
1623              **************************/
1624
1625             if (gmx_mm_any_lt(rsq21,rcutoff2))
1626             {
1627
1628             /* REACTION-FIELD ELECTROSTATICS */
1629             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1630
1631             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1632
1633             fscal            = felec;
1634
1635             fscal            = _mm_and_pd(fscal,cutoff_mask);
1636
1637             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1638
1639             /* Update vectorial force */
1640             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1641             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1642             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1643
1644             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1645             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1646             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1647
1648             }
1649
1650             /**************************
1651              * CALCULATE INTERACTIONS *
1652              **************************/
1653
1654             if (gmx_mm_any_lt(rsq22,rcutoff2))
1655             {
1656
1657             /* REACTION-FIELD ELECTROSTATICS */
1658             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1659
1660             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1661
1662             fscal            = felec;
1663
1664             fscal            = _mm_and_pd(fscal,cutoff_mask);
1665
1666             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1667
1668             /* Update vectorial force */
1669             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1670             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1671             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1672
1673             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1674             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1675             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1676
1677             }
1678
1679             /**************************
1680              * CALCULATE INTERACTIONS *
1681              **************************/
1682
1683             if (gmx_mm_any_lt(rsq23,rcutoff2))
1684             {
1685
1686             /* REACTION-FIELD ELECTROSTATICS */
1687             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1688
1689             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1690
1691             fscal            = felec;
1692
1693             fscal            = _mm_and_pd(fscal,cutoff_mask);
1694
1695             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1696
1697             /* Update vectorial force */
1698             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1699             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1700             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1701
1702             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1703             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1704             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1705
1706             }
1707
1708             /**************************
1709              * CALCULATE INTERACTIONS *
1710              **************************/
1711
1712             if (gmx_mm_any_lt(rsq31,rcutoff2))
1713             {
1714
1715             /* REACTION-FIELD ELECTROSTATICS */
1716             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1717
1718             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1719
1720             fscal            = felec;
1721
1722             fscal            = _mm_and_pd(fscal,cutoff_mask);
1723
1724             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1725
1726             /* Update vectorial force */
1727             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1728             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1729             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1730
1731             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1732             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1733             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1734
1735             }
1736
1737             /**************************
1738              * CALCULATE INTERACTIONS *
1739              **************************/
1740
1741             if (gmx_mm_any_lt(rsq32,rcutoff2))
1742             {
1743
1744             /* REACTION-FIELD ELECTROSTATICS */
1745             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1746
1747             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1748
1749             fscal            = felec;
1750
1751             fscal            = _mm_and_pd(fscal,cutoff_mask);
1752
1753             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1754
1755             /* Update vectorial force */
1756             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1757             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1758             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1759
1760             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1761             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1762             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1763
1764             }
1765
1766             /**************************
1767              * CALCULATE INTERACTIONS *
1768              **************************/
1769
1770             if (gmx_mm_any_lt(rsq33,rcutoff2))
1771             {
1772
1773             /* REACTION-FIELD ELECTROSTATICS */
1774             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1775
1776             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1777
1778             fscal            = felec;
1779
1780             fscal            = _mm_and_pd(fscal,cutoff_mask);
1781
1782             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1783
1784             /* Update vectorial force */
1785             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1786             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1787             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1788
1789             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1790             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1791             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1792
1793             }
1794
1795             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1796
1797             /* Inner loop uses 297 flops */
1798         }
1799
1800         /* End of innermost loop */
1801
1802         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1803                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1804
1805         /* Increment number of inner iterations */
1806         inneriter                  += j_index_end - j_index_start;
1807
1808         /* Outer loop uses 18 flops */
1809     }
1810
1811     /* Increment number of outer iterations */
1812     outeriter        += nri;
1813
1814     /* Update outer/inner flops */
1815
1816     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*297);
1817 }