src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_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 "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/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_VdwLJSh_GeomW3W3_VF_avx_128_fma_double
  54  * Electrostatics interaction: ReactionField
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Water3-Water3
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_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              vdwioffset0;
  83     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwioffset1;
  85     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  86     int              vdwioffset2;
  87     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  88     int              vdwjidx0A,vdwjidx0B;
  89     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     int              vdwjidx1A,vdwjidx1B;
  91     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  92     int              vdwjidx2A,vdwjidx2B;
  93     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  94     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  95     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  96     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  97     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  98     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  99     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 100     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 101     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 102     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 103     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 104     real             *charge;
 105     int              nvdwtype;
 106     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 107     int              *vdwtype;
 108     real             *vdwparam;
 109     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 110     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 111     __m128d          dummy_mask,cutoff_mask;
 112     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 113     __m128d          one     = _mm_set1_pd(1.0);
 114     __m128d          two     = _mm_set1_pd(2.0);
 115     x                = xx[0];
 116     f                = ff[0];
 117
 118     nri              = nlist->nri;
 119     iinr             = nlist->iinr;
 120     jindex           = nlist->jindex;
 121     jjnr             = nlist->jjnr;
 122     shiftidx         = nlist->shift;
 123     gid              = nlist->gid;
 124     shiftvec         = fr->shift_vec[0];
 125     fshift           = fr->fshift[0];
 126     facel            = _mm_set1_pd(fr->epsfac);
 127     charge           = mdatoms->chargeA;
 128     krf              = _mm_set1_pd(fr->ic->k_rf);
 129     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 130     crf              = _mm_set1_pd(fr->ic->c_rf);
 131     nvdwtype         = fr->ntype;
 132     vdwparam         = fr->nbfp;
 133     vdwtype          = mdatoms->typeA;
 134
 135     /* Setup water-specific parameters */
 136     inr              = nlist->iinr[0];
 137     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 138     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 139     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 140     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 141
 142     jq0              = _mm_set1_pd(charge[inr+0]);
 143     jq1              = _mm_set1_pd(charge[inr+1]);
 144     jq2              = _mm_set1_pd(charge[inr+2]);
 145     vdwjidx0A        = 2*vdwtype[inr+0];
 146     qq00             = _mm_mul_pd(iq0,jq0);
 147     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 148     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 149     qq01             = _mm_mul_pd(iq0,jq1);
 150     qq02             = _mm_mul_pd(iq0,jq2);
 151     qq10             = _mm_mul_pd(iq1,jq0);
 152     qq11             = _mm_mul_pd(iq1,jq1);
 153     qq12             = _mm_mul_pd(iq1,jq2);
 154     qq20             = _mm_mul_pd(iq2,jq0);
 155     qq21             = _mm_mul_pd(iq2,jq1);
 156     qq22             = _mm_mul_pd(iq2,jq2);
 157
 158     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 159     rcutoff_scalar   = fr->rcoulomb;
 160     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 161     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 162
 163     sh_vdw_invrcut6  = _mm_set1_pd(fr->ic->sh_invrc6);
 164     rvdw             = _mm_set1_pd(fr->rvdw);
 165
 166     /* Avoid stupid compiler warnings */
 167     jnrA = jnrB = 0;
 168     j_coord_offsetA = 0;
 169     j_coord_offsetB = 0;
 170
 171     outeriter        = 0;
 172     inneriter        = 0;
 173
 174     /* Start outer loop over neighborlists */
 175     for(iidx=0; iidx<nri; iidx++)
 176     {
 177         /* Load shift vector for this list */
 178         i_shift_offset   = DIM*shiftidx[iidx];
 179
 180         /* Load limits for loop over neighbors */
 181         j_index_start    = jindex[iidx];
 182         j_index_end      = jindex[iidx+1];
 183
 184         /* Get outer coordinate index */
 185         inr              = iinr[iidx];
 186         i_coord_offset   = DIM*inr;
 187
 188         /* Load i particle coords and add shift vector */
 189         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 190                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 191
 192         fix0             = _mm_setzero_pd();
 193         fiy0             = _mm_setzero_pd();
 194         fiz0             = _mm_setzero_pd();
 195         fix1             = _mm_setzero_pd();
 196         fiy1             = _mm_setzero_pd();
 197         fiz1             = _mm_setzero_pd();
 198         fix2             = _mm_setzero_pd();
 199         fiy2             = _mm_setzero_pd();
 200         fiz2             = _mm_setzero_pd();
 201
 202         /* Reset potential sums */
 203         velecsum         = _mm_setzero_pd();
 204         vvdwsum          = _mm_setzero_pd();
 205
 206         /* Start inner kernel loop */
 207         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 208         {
 209
 210             /* Get j neighbor index, and coordinate index */
 211             jnrA             = jjnr[jidx];
 212             jnrB             = jjnr[jidx+1];
 213             j_coord_offsetA  = DIM*jnrA;
 214             j_coord_offsetB  = DIM*jnrB;
 215
 216             /* load j atom coordinates */
 217             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 218                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 219
 220             /* Calculate displacement vector */
 221             dx00             = _mm_sub_pd(ix0,jx0);
 222             dy00             = _mm_sub_pd(iy0,jy0);
 223             dz00             = _mm_sub_pd(iz0,jz0);
 224             dx01             = _mm_sub_pd(ix0,jx1);
 225             dy01             = _mm_sub_pd(iy0,jy1);
 226             dz01             = _mm_sub_pd(iz0,jz1);
 227             dx02             = _mm_sub_pd(ix0,jx2);
 228             dy02             = _mm_sub_pd(iy0,jy2);
 229             dz02             = _mm_sub_pd(iz0,jz2);
 230             dx10             = _mm_sub_pd(ix1,jx0);
 231             dy10             = _mm_sub_pd(iy1,jy0);
 232             dz10             = _mm_sub_pd(iz1,jz0);
 233             dx11             = _mm_sub_pd(ix1,jx1);
 234             dy11             = _mm_sub_pd(iy1,jy1);
 235             dz11             = _mm_sub_pd(iz1,jz1);
 236             dx12             = _mm_sub_pd(ix1,jx2);
 237             dy12             = _mm_sub_pd(iy1,jy2);
 238             dz12             = _mm_sub_pd(iz1,jz2);
 239             dx20             = _mm_sub_pd(ix2,jx0);
 240             dy20             = _mm_sub_pd(iy2,jy0);
 241             dz20             = _mm_sub_pd(iz2,jz0);
 242             dx21             = _mm_sub_pd(ix2,jx1);
 243             dy21             = _mm_sub_pd(iy2,jy1);
 244             dz21             = _mm_sub_pd(iz2,jz1);
 245             dx22             = _mm_sub_pd(ix2,jx2);
 246             dy22             = _mm_sub_pd(iy2,jy2);
 247             dz22             = _mm_sub_pd(iz2,jz2);
 248
 249             /* Calculate squared distance and things based on it */
 250             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 251             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 252             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 253             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 254             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 255             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 256             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 257             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 258             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 259
 260             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 261             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 262             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 263             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 264             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 265             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 266             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 267             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 268             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 269
 270             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 271             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 272             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 273             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 274             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 275             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 276             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 277             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 278             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 279
 280             fjx0             = _mm_setzero_pd();
 281             fjy0             = _mm_setzero_pd();
 282             fjz0             = _mm_setzero_pd();
 283             fjx1             = _mm_setzero_pd();
 284             fjy1             = _mm_setzero_pd();
 285             fjz1             = _mm_setzero_pd();
 286             fjx2             = _mm_setzero_pd();
 287             fjy2             = _mm_setzero_pd();
 288             fjz2             = _mm_setzero_pd();
 289
 290             /**************************
 291              * CALCULATE INTERACTIONS *
 292              **************************/
 293
 294             if (gmx_mm_any_lt(rsq00,rcutoff2))
 295             {
 296
 297             /* REACTION-FIELD ELECTROSTATICS */
 298             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
 299             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 300
 301             /* LENNARD-JONES DISPERSION/REPULSION */
 302
 303             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 304             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 305             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 306             vvdw             = _mm_msub_pd(_mm_nmacc_pd(c12_00,_mm_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 307                                            _mm_mul_pd(_mm_nmacc_pd( c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 308             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 309
 310             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 311
 312             /* Update potential sum for this i atom from the interaction with this j atom. */
 313             velec            = _mm_and_pd(velec,cutoff_mask);
 314             velecsum         = _mm_add_pd(velecsum,velec);
 315             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 316             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 317
 318             fscal            = _mm_add_pd(felec,fvdw);
 319
 320             fscal            = _mm_and_pd(fscal,cutoff_mask);
 321
 322             /* Update vectorial force */
 323             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 324             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 325             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 326
 327             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 328             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 329             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 330
 331             }
 332
 333             /**************************
 334              * CALCULATE INTERACTIONS *
 335              **************************/
 336
 337             if (gmx_mm_any_lt(rsq01,rcutoff2))
 338             {
 339
 340             /* REACTION-FIELD ELECTROSTATICS */
 341             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_macc_pd(krf,rsq01,rinv01),crf));
 342             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
 343
 344             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
 345
 346             /* Update potential sum for this i atom from the interaction with this j atom. */
 347             velec            = _mm_and_pd(velec,cutoff_mask);
 348             velecsum         = _mm_add_pd(velecsum,velec);
 349
 350             fscal            = felec;
 351
 352             fscal            = _mm_and_pd(fscal,cutoff_mask);
 353
 354             /* Update vectorial force */
 355             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 356             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 357             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 358
 359             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 360             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 361             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 362
 363             }
 364
 365             /**************************
 366              * CALCULATE INTERACTIONS *
 367              **************************/
 368
 369             if (gmx_mm_any_lt(rsq02,rcutoff2))
 370             {
 371
 372             /* REACTION-FIELD ELECTROSTATICS */
 373             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_macc_pd(krf,rsq02,rinv02),crf));
 374             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
 375
 376             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
 377
 378             /* Update potential sum for this i atom from the interaction with this j atom. */
 379             velec            = _mm_and_pd(velec,cutoff_mask);
 380             velecsum         = _mm_add_pd(velecsum,velec);
 381
 382             fscal            = felec;
 383
 384             fscal            = _mm_and_pd(fscal,cutoff_mask);
 385
 386             /* Update vectorial force */
 387             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 388             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 389             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 390
 391             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 392             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 393             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 394
 395             }
 396
 397             /**************************
 398              * CALCULATE INTERACTIONS *
 399              **************************/
 400
 401             if (gmx_mm_any_lt(rsq10,rcutoff2))
 402             {
 403
 404             /* REACTION-FIELD ELECTROSTATICS */
 405             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
 406             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 407
 408             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
 409
 410             /* Update potential sum for this i atom from the interaction with this j atom. */
 411             velec            = _mm_and_pd(velec,cutoff_mask);
 412             velecsum         = _mm_add_pd(velecsum,velec);
 413
 414             fscal            = felec;
 415
 416             fscal            = _mm_and_pd(fscal,cutoff_mask);
 417
 418             /* Update vectorial force */
 419             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 420             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 421             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 422
 423             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 424             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 425             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 426
 427             }
 428
 429             /**************************
 430              * CALCULATE INTERACTIONS *
 431              **************************/
 432
 433             if (gmx_mm_any_lt(rsq11,rcutoff2))
 434             {
 435
 436             /* REACTION-FIELD ELECTROSTATICS */
 437             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 438             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 439
 440             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 441
 442             /* Update potential sum for this i atom from the interaction with this j atom. */
 443             velec            = _mm_and_pd(velec,cutoff_mask);
 444             velecsum         = _mm_add_pd(velecsum,velec);
 445
 446             fscal            = felec;
 447
 448             fscal            = _mm_and_pd(fscal,cutoff_mask);
 449
 450             /* Update vectorial force */
 451             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 452             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 453             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 454
 455             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 456             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 457             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 458
 459             }
 460
 461             /**************************
 462              * CALCULATE INTERACTIONS *
 463              **************************/
 464
 465             if (gmx_mm_any_lt(rsq12,rcutoff2))
 466             {
 467
 468             /* REACTION-FIELD ELECTROSTATICS */
 469             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 470             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 471
 472             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 473
 474             /* Update potential sum for this i atom from the interaction with this j atom. */
 475             velec            = _mm_and_pd(velec,cutoff_mask);
 476             velecsum         = _mm_add_pd(velecsum,velec);
 477
 478             fscal            = felec;
 479
 480             fscal            = _mm_and_pd(fscal,cutoff_mask);
 481
 482             /* Update vectorial force */
 483             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 484             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 485             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 486
 487             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 488             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 489             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 490
 491             }
 492
 493             /**************************
 494              * CALCULATE INTERACTIONS *
 495              **************************/
 496
 497             if (gmx_mm_any_lt(rsq20,rcutoff2))
 498             {
 499
 500             /* REACTION-FIELD ELECTROSTATICS */
 501             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
 502             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 503
 504             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
 505
 506             /* Update potential sum for this i atom from the interaction with this j atom. */
 507             velec            = _mm_and_pd(velec,cutoff_mask);
 508             velecsum         = _mm_add_pd(velecsum,velec);
 509
 510             fscal            = felec;
 511
 512             fscal            = _mm_and_pd(fscal,cutoff_mask);
 513
 514             /* Update vectorial force */
 515             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 516             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 517             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 518
 519             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 520             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 521             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 522
 523             }
 524
 525             /**************************
 526              * CALCULATE INTERACTIONS *
 527              **************************/
 528
 529             if (gmx_mm_any_lt(rsq21,rcutoff2))
 530             {
 531
 532             /* REACTION-FIELD ELECTROSTATICS */
 533             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 534             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 535
 536             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 537
 538             /* Update potential sum for this i atom from the interaction with this j atom. */
 539             velec            = _mm_and_pd(velec,cutoff_mask);
 540             velecsum         = _mm_add_pd(velecsum,velec);
 541
 542             fscal            = felec;
 543
 544             fscal            = _mm_and_pd(fscal,cutoff_mask);
 545
 546             /* Update vectorial force */
 547             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 548             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 549             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 550
 551             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 552             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 553             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 554
 555             }
 556
 557             /**************************
 558              * CALCULATE INTERACTIONS *
 559              **************************/
 560
 561             if (gmx_mm_any_lt(rsq22,rcutoff2))
 562             {
 563
 564             /* REACTION-FIELD ELECTROSTATICS */
 565             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 566             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 567
 568             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 569
 570             /* Update potential sum for this i atom from the interaction with this j atom. */
 571             velec            = _mm_and_pd(velec,cutoff_mask);
 572             velecsum         = _mm_add_pd(velecsum,velec);
 573
 574             fscal            = felec;
 575
 576             fscal            = _mm_and_pd(fscal,cutoff_mask);
 577
 578             /* Update vectorial force */
 579             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 580             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 581             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 582
 583             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 584             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 585             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 586
 587             }
 588
 589             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 590
 591             /* Inner loop uses 369 flops */
 592         }
 593
 594         if(jidx<j_index_end)
 595         {
 596
 597             jnrA             = jjnr[jidx];
 598             j_coord_offsetA  = DIM*jnrA;
 599
 600             /* load j atom coordinates */
 601             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 602                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 603
 604             /* Calculate displacement vector */
 605             dx00             = _mm_sub_pd(ix0,jx0);
 606             dy00             = _mm_sub_pd(iy0,jy0);
 607             dz00             = _mm_sub_pd(iz0,jz0);
 608             dx01             = _mm_sub_pd(ix0,jx1);
 609             dy01             = _mm_sub_pd(iy0,jy1);
 610             dz01             = _mm_sub_pd(iz0,jz1);
 611             dx02             = _mm_sub_pd(ix0,jx2);
 612             dy02             = _mm_sub_pd(iy0,jy2);
 613             dz02             = _mm_sub_pd(iz0,jz2);
 614             dx10             = _mm_sub_pd(ix1,jx0);
 615             dy10             = _mm_sub_pd(iy1,jy0);
 616             dz10             = _mm_sub_pd(iz1,jz0);
 617             dx11             = _mm_sub_pd(ix1,jx1);
 618             dy11             = _mm_sub_pd(iy1,jy1);
 619             dz11             = _mm_sub_pd(iz1,jz1);
 620             dx12             = _mm_sub_pd(ix1,jx2);
 621             dy12             = _mm_sub_pd(iy1,jy2);
 622             dz12             = _mm_sub_pd(iz1,jz2);
 623             dx20             = _mm_sub_pd(ix2,jx0);
 624             dy20             = _mm_sub_pd(iy2,jy0);
 625             dz20             = _mm_sub_pd(iz2,jz0);
 626             dx21             = _mm_sub_pd(ix2,jx1);
 627             dy21             = _mm_sub_pd(iy2,jy1);
 628             dz21             = _mm_sub_pd(iz2,jz1);
 629             dx22             = _mm_sub_pd(ix2,jx2);
 630             dy22             = _mm_sub_pd(iy2,jy2);
 631             dz22             = _mm_sub_pd(iz2,jz2);
 632
 633             /* Calculate squared distance and things based on it */
 634             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 635             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 636             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 637             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 638             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 639             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 640             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 641             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 642             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 643
 644             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 645             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 646             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 647             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 648             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 649             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 650             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 651             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 652             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 653
 654             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 655             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 656             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 657             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 658             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 659             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 660             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 661             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 662             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 663
 664             fjx0             = _mm_setzero_pd();
 665             fjy0             = _mm_setzero_pd();
 666             fjz0             = _mm_setzero_pd();
 667             fjx1             = _mm_setzero_pd();
 668             fjy1             = _mm_setzero_pd();
 669             fjz1             = _mm_setzero_pd();
 670             fjx2             = _mm_setzero_pd();
 671             fjy2             = _mm_setzero_pd();
 672             fjz2             = _mm_setzero_pd();
 673
 674             /**************************
 675              * CALCULATE INTERACTIONS *
 676              **************************/
 677
 678             if (gmx_mm_any_lt(rsq00,rcutoff2))
 679             {
 680
 681             /* REACTION-FIELD ELECTROSTATICS */
 682             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
 683             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 684
 685             /* LENNARD-JONES DISPERSION/REPULSION */
 686
 687             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 688             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 689             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 690             vvdw             = _mm_msub_pd(_mm_nmacc_pd(c12_00,_mm_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 691                                            _mm_mul_pd(_mm_nmacc_pd( c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 692             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 693
 694             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 695
 696             /* Update potential sum for this i atom from the interaction with this j atom. */
 697             velec            = _mm_and_pd(velec,cutoff_mask);
 698             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 699             velecsum         = _mm_add_pd(velecsum,velec);
 700             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 701             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 702             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 703
 704             fscal            = _mm_add_pd(felec,fvdw);
 705
 706             fscal            = _mm_and_pd(fscal,cutoff_mask);
 707
 708             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 709
 710             /* Update vectorial force */
 711             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 712             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 713             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 714
 715             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 716             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 717             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 718
 719             }
 720
 721             /**************************
 722              * CALCULATE INTERACTIONS *
 723              **************************/
 724
 725             if (gmx_mm_any_lt(rsq01,rcutoff2))
 726             {
 727
 728             /* REACTION-FIELD ELECTROSTATICS */
 729             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_macc_pd(krf,rsq01,rinv01),crf));
 730             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
 731
 732             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
 733
 734             /* Update potential sum for this i atom from the interaction with this j atom. */
 735             velec            = _mm_and_pd(velec,cutoff_mask);
 736             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 737             velecsum         = _mm_add_pd(velecsum,velec);
 738
 739             fscal            = felec;
 740
 741             fscal            = _mm_and_pd(fscal,cutoff_mask);
 742
 743             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 744
 745             /* Update vectorial force */
 746             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 747             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 748             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 749
 750             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 751             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 752             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 753
 754             }
 755
 756             /**************************
 757              * CALCULATE INTERACTIONS *
 758              **************************/
 759
 760             if (gmx_mm_any_lt(rsq02,rcutoff2))
 761             {
 762
 763             /* REACTION-FIELD ELECTROSTATICS */
 764             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_macc_pd(krf,rsq02,rinv02),crf));
 765             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
 766
 767             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
 768
 769             /* Update potential sum for this i atom from the interaction with this j atom. */
 770             velec            = _mm_and_pd(velec,cutoff_mask);
 771             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 772             velecsum         = _mm_add_pd(velecsum,velec);
 773
 774             fscal            = felec;
 775
 776             fscal            = _mm_and_pd(fscal,cutoff_mask);
 777
 778             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 779
 780             /* Update vectorial force */
 781             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 782             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 783             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 784
 785             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 786             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 787             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 788
 789             }
 790
 791             /**************************
 792              * CALCULATE INTERACTIONS *
 793              **************************/
 794
 795             if (gmx_mm_any_lt(rsq10,rcutoff2))
 796             {
 797
 798             /* REACTION-FIELD ELECTROSTATICS */
 799             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
 800             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 801
 802             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
 803
 804             /* Update potential sum for this i atom from the interaction with this j atom. */
 805             velec            = _mm_and_pd(velec,cutoff_mask);
 806             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 807             velecsum         = _mm_add_pd(velecsum,velec);
 808
 809             fscal            = felec;
 810
 811             fscal            = _mm_and_pd(fscal,cutoff_mask);
 812
 813             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 814
 815             /* Update vectorial force */
 816             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 817             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 818             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 819
 820             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 821             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 822             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 823
 824             }
 825
 826             /**************************
 827              * CALCULATE INTERACTIONS *
 828              **************************/
 829
 830             if (gmx_mm_any_lt(rsq11,rcutoff2))
 831             {
 832
 833             /* REACTION-FIELD ELECTROSTATICS */
 834             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 835             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 836
 837             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 838
 839             /* Update potential sum for this i atom from the interaction with this j atom. */
 840             velec            = _mm_and_pd(velec,cutoff_mask);
 841             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 842             velecsum         = _mm_add_pd(velecsum,velec);
 843
 844             fscal            = felec;
 845
 846             fscal            = _mm_and_pd(fscal,cutoff_mask);
 847
 848             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 849
 850             /* Update vectorial force */
 851             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 852             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 853             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 854
 855             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 856             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 857             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 858
 859             }
 860
 861             /**************************
 862              * CALCULATE INTERACTIONS *
 863              **************************/
 864
 865             if (gmx_mm_any_lt(rsq12,rcutoff2))
 866             {
 867
 868             /* REACTION-FIELD ELECTROSTATICS */
 869             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 870             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 871
 872             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 873
 874             /* Update potential sum for this i atom from the interaction with this j atom. */
 875             velec            = _mm_and_pd(velec,cutoff_mask);
 876             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 877             velecsum         = _mm_add_pd(velecsum,velec);
 878
 879             fscal            = felec;
 880
 881             fscal            = _mm_and_pd(fscal,cutoff_mask);
 882
 883             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 884
 885             /* Update vectorial force */
 886             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 887             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 888             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 889
 890             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 891             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 892             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 893
 894             }
 895
 896             /**************************
 897              * CALCULATE INTERACTIONS *
 898              **************************/
 899
 900             if (gmx_mm_any_lt(rsq20,rcutoff2))
 901             {
 902
 903             /* REACTION-FIELD ELECTROSTATICS */
 904             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
 905             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 906
 907             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
 908
 909             /* Update potential sum for this i atom from the interaction with this j atom. */
 910             velec            = _mm_and_pd(velec,cutoff_mask);
 911             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 912             velecsum         = _mm_add_pd(velecsum,velec);
 913
 914             fscal            = felec;
 915
 916             fscal            = _mm_and_pd(fscal,cutoff_mask);
 917
 918             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 919
 920             /* Update vectorial force */
 921             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 922             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 923             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 924
 925             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 926             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 927             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 928
 929             }
 930
 931             /**************************
 932              * CALCULATE INTERACTIONS *
 933              **************************/
 934
 935             if (gmx_mm_any_lt(rsq21,rcutoff2))
 936             {
 937
 938             /* REACTION-FIELD ELECTROSTATICS */
 939             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 940             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 941
 942             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 943
 944             /* Update potential sum for this i atom from the interaction with this j atom. */
 945             velec            = _mm_and_pd(velec,cutoff_mask);
 946             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 947             velecsum         = _mm_add_pd(velecsum,velec);
 948
 949             fscal            = felec;
 950
 951             fscal            = _mm_and_pd(fscal,cutoff_mask);
 952
 953             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 954
 955             /* Update vectorial force */
 956             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 957             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 958             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 959
 960             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 961             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 962             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 963
 964             }
 965
 966             /**************************
 967              * CALCULATE INTERACTIONS *
 968              **************************/
 969
 970             if (gmx_mm_any_lt(rsq22,rcutoff2))
 971             {
 972
 973             /* REACTION-FIELD ELECTROSTATICS */
 974             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 975             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 976
 977             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 978
 979             /* Update potential sum for this i atom from the interaction with this j atom. */
 980             velec            = _mm_and_pd(velec,cutoff_mask);
 981             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 982             velecsum         = _mm_add_pd(velecsum,velec);
 983
 984             fscal            = felec;
 985
 986             fscal            = _mm_and_pd(fscal,cutoff_mask);
 987
 988             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 989
 990             /* Update vectorial force */
 991             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 992             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 993             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 994
 995             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 996             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 997             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 998
 999             }
1000
1001             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1002
1003             /* Inner loop uses 369 flops */
1004         }
1005
1006         /* End of innermost loop */
1007
1008         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1009                                               f+i_coord_offset,fshift+i_shift_offset);
1010
1011         ggid                        = gid[iidx];
1012         /* Update potential energies */
1013         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1014         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
1015
1016         /* Increment number of inner iterations */
1017         inneriter                  += j_index_end - j_index_start;
1018
1019         /* Outer loop uses 20 flops */
1020     }
1021
1022     /* Increment number of outer iterations */
1023     outeriter        += nri;
1024
1025     /* Update outer/inner flops */
1026
1027     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*369);
1028 }
1029 /*
1030  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_F_avx_128_fma_double
1031  * Electrostatics interaction: ReactionField
1032  * VdW interaction:            LennardJones
1033  * Geometry:                   Water3-Water3
1034  * Calculate force/pot:        Force
1035  */
1036 void
1037 nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_F_avx_128_fma_double
1038                     (t_nblist                    * gmx_restrict       nlist,
1039                      rvec                        * gmx_restrict          xx,
1040                      rvec                        * gmx_restrict          ff,
1041                      t_forcerec                  * gmx_restrict          fr,
1042                      t_mdatoms                   * gmx_restrict     mdatoms,
1043                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1044                      t_nrnb                      * gmx_restrict        nrnb)
1045 {
1046     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1047      * just 0 for non-waters.
1048      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1049      * jnr indices corresponding to data put in the four positions in the SIMD register.
1050      */
1051     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1052     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1053     int              jnrA,jnrB;
1054     int              j_coord_offsetA,j_coord_offsetB;
1055     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1056     real             rcutoff_scalar;
1057     real             *shiftvec,*fshift,*x,*f;
1058     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1059     int              vdwioffset0;
1060     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1061     int              vdwioffset1;
1062     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1063     int              vdwioffset2;
1064     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1065     int              vdwjidx0A,vdwjidx0B;
1066     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1067     int              vdwjidx1A,vdwjidx1B;
1068     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1069     int              vdwjidx2A,vdwjidx2B;
1070     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1071     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1072     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1073     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1074     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1075     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1076     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1077     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1078     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1079     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1080     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1081     real             *charge;
1082     int              nvdwtype;
1083     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1084     int              *vdwtype;
1085     real             *vdwparam;
1086     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1087     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1088     __m128d          dummy_mask,cutoff_mask;
1089     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1090     __m128d          one     = _mm_set1_pd(1.0);
1091     __m128d          two     = _mm_set1_pd(2.0);
1092     x                = xx[0];
1093     f                = ff[0];
1094
1095     nri              = nlist->nri;
1096     iinr             = nlist->iinr;
1097     jindex           = nlist->jindex;
1098     jjnr             = nlist->jjnr;
1099     shiftidx         = nlist->shift;
1100     gid              = nlist->gid;
1101     shiftvec         = fr->shift_vec[0];
1102     fshift           = fr->fshift[0];
1103     facel            = _mm_set1_pd(fr->epsfac);
1104     charge           = mdatoms->chargeA;
1105     krf              = _mm_set1_pd(fr->ic->k_rf);
1106     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1107     crf              = _mm_set1_pd(fr->ic->c_rf);
1108     nvdwtype         = fr->ntype;
1109     vdwparam         = fr->nbfp;
1110     vdwtype          = mdatoms->typeA;
1111
1112     /* Setup water-specific parameters */
1113     inr              = nlist->iinr[0];
1114     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
1115     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1116     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1117     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1118
1119     jq0              = _mm_set1_pd(charge[inr+0]);
1120     jq1              = _mm_set1_pd(charge[inr+1]);
1121     jq2              = _mm_set1_pd(charge[inr+2]);
1122     vdwjidx0A        = 2*vdwtype[inr+0];
1123     qq00             = _mm_mul_pd(iq0,jq0);
1124     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1125     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1126     qq01             = _mm_mul_pd(iq0,jq1);
1127     qq02             = _mm_mul_pd(iq0,jq2);
1128     qq10             = _mm_mul_pd(iq1,jq0);
1129     qq11             = _mm_mul_pd(iq1,jq1);
1130     qq12             = _mm_mul_pd(iq1,jq2);
1131     qq20             = _mm_mul_pd(iq2,jq0);
1132     qq21             = _mm_mul_pd(iq2,jq1);
1133     qq22             = _mm_mul_pd(iq2,jq2);
1134
1135     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1136     rcutoff_scalar   = fr->rcoulomb;
1137     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1138     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1139
1140     sh_vdw_invrcut6  = _mm_set1_pd(fr->ic->sh_invrc6);
1141     rvdw             = _mm_set1_pd(fr->rvdw);
1142
1143     /* Avoid stupid compiler warnings */
1144     jnrA = jnrB = 0;
1145     j_coord_offsetA = 0;
1146     j_coord_offsetB = 0;
1147
1148     outeriter        = 0;
1149     inneriter        = 0;
1150
1151     /* Start outer loop over neighborlists */
1152     for(iidx=0; iidx<nri; iidx++)
1153     {
1154         /* Load shift vector for this list */
1155         i_shift_offset   = DIM*shiftidx[iidx];
1156
1157         /* Load limits for loop over neighbors */
1158         j_index_start    = jindex[iidx];
1159         j_index_end      = jindex[iidx+1];
1160
1161         /* Get outer coordinate index */
1162         inr              = iinr[iidx];
1163         i_coord_offset   = DIM*inr;
1164
1165         /* Load i particle coords and add shift vector */
1166         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1167                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1168
1169         fix0             = _mm_setzero_pd();
1170         fiy0             = _mm_setzero_pd();
1171         fiz0             = _mm_setzero_pd();
1172         fix1             = _mm_setzero_pd();
1173         fiy1             = _mm_setzero_pd();
1174         fiz1             = _mm_setzero_pd();
1175         fix2             = _mm_setzero_pd();
1176         fiy2             = _mm_setzero_pd();
1177         fiz2             = _mm_setzero_pd();
1178
1179         /* Start inner kernel loop */
1180         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1181         {
1182
1183             /* Get j neighbor index, and coordinate index */
1184             jnrA             = jjnr[jidx];
1185             jnrB             = jjnr[jidx+1];
1186             j_coord_offsetA  = DIM*jnrA;
1187             j_coord_offsetB  = DIM*jnrB;
1188
1189             /* load j atom coordinates */
1190             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1191                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1192
1193             /* Calculate displacement vector */
1194             dx00             = _mm_sub_pd(ix0,jx0);
1195             dy00             = _mm_sub_pd(iy0,jy0);
1196             dz00             = _mm_sub_pd(iz0,jz0);
1197             dx01             = _mm_sub_pd(ix0,jx1);
1198             dy01             = _mm_sub_pd(iy0,jy1);
1199             dz01             = _mm_sub_pd(iz0,jz1);
1200             dx02             = _mm_sub_pd(ix0,jx2);
1201             dy02             = _mm_sub_pd(iy0,jy2);
1202             dz02             = _mm_sub_pd(iz0,jz2);
1203             dx10             = _mm_sub_pd(ix1,jx0);
1204             dy10             = _mm_sub_pd(iy1,jy0);
1205             dz10             = _mm_sub_pd(iz1,jz0);
1206             dx11             = _mm_sub_pd(ix1,jx1);
1207             dy11             = _mm_sub_pd(iy1,jy1);
1208             dz11             = _mm_sub_pd(iz1,jz1);
1209             dx12             = _mm_sub_pd(ix1,jx2);
1210             dy12             = _mm_sub_pd(iy1,jy2);
1211             dz12             = _mm_sub_pd(iz1,jz2);
1212             dx20             = _mm_sub_pd(ix2,jx0);
1213             dy20             = _mm_sub_pd(iy2,jy0);
1214             dz20             = _mm_sub_pd(iz2,jz0);
1215             dx21             = _mm_sub_pd(ix2,jx1);
1216             dy21             = _mm_sub_pd(iy2,jy1);
1217             dz21             = _mm_sub_pd(iz2,jz1);
1218             dx22             = _mm_sub_pd(ix2,jx2);
1219             dy22             = _mm_sub_pd(iy2,jy2);
1220             dz22             = _mm_sub_pd(iz2,jz2);
1221
1222             /* Calculate squared distance and things based on it */
1223             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1224             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1225             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1226             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1227             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1228             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1229             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1230             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1231             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1232
1233             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1234             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1235             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1236             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1237             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1238             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1239             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1240             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1241             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1242
1243             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1244             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1245             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1246             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1247             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1248             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1249             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1250             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1251             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1252
1253             fjx0             = _mm_setzero_pd();
1254             fjy0             = _mm_setzero_pd();
1255             fjz0             = _mm_setzero_pd();
1256             fjx1             = _mm_setzero_pd();
1257             fjy1             = _mm_setzero_pd();
1258             fjz1             = _mm_setzero_pd();
1259             fjx2             = _mm_setzero_pd();
1260             fjy2             = _mm_setzero_pd();
1261             fjz2             = _mm_setzero_pd();
1262
1263             /**************************
1264              * CALCULATE INTERACTIONS *
1265              **************************/
1266
1267             if (gmx_mm_any_lt(rsq00,rcutoff2))
1268             {
1269
1270             /* REACTION-FIELD ELECTROSTATICS */
1271             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
1272
1273             /* LENNARD-JONES DISPERSION/REPULSION */
1274
1275             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1276             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1277
1278             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1279
1280             fscal            = _mm_add_pd(felec,fvdw);
1281
1282             fscal            = _mm_and_pd(fscal,cutoff_mask);
1283
1284             /* Update vectorial force */
1285             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1286             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1287             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1288
1289             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1290             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1291             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1292
1293             }
1294
1295             /**************************
1296              * CALCULATE INTERACTIONS *
1297              **************************/
1298
1299             if (gmx_mm_any_lt(rsq01,rcutoff2))
1300             {
1301
1302             /* REACTION-FIELD ELECTROSTATICS */
1303             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
1304
1305             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
1306
1307             fscal            = felec;
1308
1309             fscal            = _mm_and_pd(fscal,cutoff_mask);
1310
1311             /* Update vectorial force */
1312             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1313             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1314             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1315
1316             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1317             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1318             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1319
1320             }
1321
1322             /**************************
1323              * CALCULATE INTERACTIONS *
1324              **************************/
1325
1326             if (gmx_mm_any_lt(rsq02,rcutoff2))
1327             {
1328
1329             /* REACTION-FIELD ELECTROSTATICS */
1330             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
1331
1332             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
1333
1334             fscal            = felec;
1335
1336             fscal            = _mm_and_pd(fscal,cutoff_mask);
1337
1338             /* Update vectorial force */
1339             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1340             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1341             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1342
1343             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1344             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1345             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1346
1347             }
1348
1349             /**************************
1350              * CALCULATE INTERACTIONS *
1351              **************************/
1352
1353             if (gmx_mm_any_lt(rsq10,rcutoff2))
1354             {
1355
1356             /* REACTION-FIELD ELECTROSTATICS */
1357             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
1358
1359             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
1360
1361             fscal            = felec;
1362
1363             fscal            = _mm_and_pd(fscal,cutoff_mask);
1364
1365             /* Update vectorial force */
1366             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1367             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1368             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1369
1370             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1371             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1372             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1373
1374             }
1375
1376             /**************************
1377              * CALCULATE INTERACTIONS *
1378              **************************/
1379
1380             if (gmx_mm_any_lt(rsq11,rcutoff2))
1381             {
1382
1383             /* REACTION-FIELD ELECTROSTATICS */
1384             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1385
1386             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1387
1388             fscal            = felec;
1389
1390             fscal            = _mm_and_pd(fscal,cutoff_mask);
1391
1392             /* Update vectorial force */
1393             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1394             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1395             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1396
1397             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1398             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1399             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1400
1401             }
1402
1403             /**************************
1404              * CALCULATE INTERACTIONS *
1405              **************************/
1406
1407             if (gmx_mm_any_lt(rsq12,rcutoff2))
1408             {
1409
1410             /* REACTION-FIELD ELECTROSTATICS */
1411             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1412
1413             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1414
1415             fscal            = felec;
1416
1417             fscal            = _mm_and_pd(fscal,cutoff_mask);
1418
1419             /* Update vectorial force */
1420             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1421             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1422             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1423
1424             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1425             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1426             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1427
1428             }
1429
1430             /**************************
1431              * CALCULATE INTERACTIONS *
1432              **************************/
1433
1434             if (gmx_mm_any_lt(rsq20,rcutoff2))
1435             {
1436
1437             /* REACTION-FIELD ELECTROSTATICS */
1438             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
1439
1440             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
1441
1442             fscal            = felec;
1443
1444             fscal            = _mm_and_pd(fscal,cutoff_mask);
1445
1446             /* Update vectorial force */
1447             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1448             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1449             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1450
1451             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1452             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1453             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1454
1455             }
1456
1457             /**************************
1458              * CALCULATE INTERACTIONS *
1459              **************************/
1460
1461             if (gmx_mm_any_lt(rsq21,rcutoff2))
1462             {
1463
1464             /* REACTION-FIELD ELECTROSTATICS */
1465             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1466
1467             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1468
1469             fscal            = felec;
1470
1471             fscal            = _mm_and_pd(fscal,cutoff_mask);
1472
1473             /* Update vectorial force */
1474             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1475             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1476             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1477
1478             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1479             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1480             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1481
1482             }
1483
1484             /**************************
1485              * CALCULATE INTERACTIONS *
1486              **************************/
1487
1488             if (gmx_mm_any_lt(rsq22,rcutoff2))
1489             {
1490
1491             /* REACTION-FIELD ELECTROSTATICS */
1492             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1493
1494             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1495
1496             fscal            = felec;
1497
1498             fscal            = _mm_and_pd(fscal,cutoff_mask);
1499
1500             /* Update vectorial force */
1501             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1502             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1503             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1504
1505             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1506             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1507             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1508
1509             }
1510
1511             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1512
1513             /* Inner loop uses 304 flops */
1514         }
1515
1516         if(jidx<j_index_end)
1517         {
1518
1519             jnrA             = jjnr[jidx];
1520             j_coord_offsetA  = DIM*jnrA;
1521
1522             /* load j atom coordinates */
1523             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1524                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1525
1526             /* Calculate displacement vector */
1527             dx00             = _mm_sub_pd(ix0,jx0);
1528             dy00             = _mm_sub_pd(iy0,jy0);
1529             dz00             = _mm_sub_pd(iz0,jz0);
1530             dx01             = _mm_sub_pd(ix0,jx1);
1531             dy01             = _mm_sub_pd(iy0,jy1);
1532             dz01             = _mm_sub_pd(iz0,jz1);
1533             dx02             = _mm_sub_pd(ix0,jx2);
1534             dy02             = _mm_sub_pd(iy0,jy2);
1535             dz02             = _mm_sub_pd(iz0,jz2);
1536             dx10             = _mm_sub_pd(ix1,jx0);
1537             dy10             = _mm_sub_pd(iy1,jy0);
1538             dz10             = _mm_sub_pd(iz1,jz0);
1539             dx11             = _mm_sub_pd(ix1,jx1);
1540             dy11             = _mm_sub_pd(iy1,jy1);
1541             dz11             = _mm_sub_pd(iz1,jz1);
1542             dx12             = _mm_sub_pd(ix1,jx2);
1543             dy12             = _mm_sub_pd(iy1,jy2);
1544             dz12             = _mm_sub_pd(iz1,jz2);
1545             dx20             = _mm_sub_pd(ix2,jx0);
1546             dy20             = _mm_sub_pd(iy2,jy0);
1547             dz20             = _mm_sub_pd(iz2,jz0);
1548             dx21             = _mm_sub_pd(ix2,jx1);
1549             dy21             = _mm_sub_pd(iy2,jy1);
1550             dz21             = _mm_sub_pd(iz2,jz1);
1551             dx22             = _mm_sub_pd(ix2,jx2);
1552             dy22             = _mm_sub_pd(iy2,jy2);
1553             dz22             = _mm_sub_pd(iz2,jz2);
1554
1555             /* Calculate squared distance and things based on it */
1556             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1557             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1558             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1559             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1560             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1561             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1562             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1563             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1564             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1565
1566             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1567             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1568             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1569             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1570             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1571             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1572             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1573             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1574             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1575
1576             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1577             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1578             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1579             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1580             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1581             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1582             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1583             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1584             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1585
1586             fjx0             = _mm_setzero_pd();
1587             fjy0             = _mm_setzero_pd();
1588             fjz0             = _mm_setzero_pd();
1589             fjx1             = _mm_setzero_pd();
1590             fjy1             = _mm_setzero_pd();
1591             fjz1             = _mm_setzero_pd();
1592             fjx2             = _mm_setzero_pd();
1593             fjy2             = _mm_setzero_pd();
1594             fjz2             = _mm_setzero_pd();
1595
1596             /**************************
1597              * CALCULATE INTERACTIONS *
1598              **************************/
1599
1600             if (gmx_mm_any_lt(rsq00,rcutoff2))
1601             {
1602
1603             /* REACTION-FIELD ELECTROSTATICS */
1604             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
1605
1606             /* LENNARD-JONES DISPERSION/REPULSION */
1607
1608             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1609             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1610
1611             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1612
1613             fscal            = _mm_add_pd(felec,fvdw);
1614
1615             fscal            = _mm_and_pd(fscal,cutoff_mask);
1616
1617             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1618
1619             /* Update vectorial force */
1620             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1621             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1622             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1623
1624             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1625             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1626             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1627
1628             }
1629
1630             /**************************
1631              * CALCULATE INTERACTIONS *
1632              **************************/
1633
1634             if (gmx_mm_any_lt(rsq01,rcutoff2))
1635             {
1636
1637             /* REACTION-FIELD ELECTROSTATICS */
1638             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
1639
1640             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
1641
1642             fscal            = felec;
1643
1644             fscal            = _mm_and_pd(fscal,cutoff_mask);
1645
1646             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1647
1648             /* Update vectorial force */
1649             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1650             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1651             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1652
1653             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1654             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1655             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1656
1657             }
1658
1659             /**************************
1660              * CALCULATE INTERACTIONS *
1661              **************************/
1662
1663             if (gmx_mm_any_lt(rsq02,rcutoff2))
1664             {
1665
1666             /* REACTION-FIELD ELECTROSTATICS */
1667             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
1668
1669             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
1670
1671             fscal            = felec;
1672
1673             fscal            = _mm_and_pd(fscal,cutoff_mask);
1674
1675             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1676
1677             /* Update vectorial force */
1678             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1679             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1680             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1681
1682             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1683             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1684             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1685
1686             }
1687
1688             /**************************
1689              * CALCULATE INTERACTIONS *
1690              **************************/
1691
1692             if (gmx_mm_any_lt(rsq10,rcutoff2))
1693             {
1694
1695             /* REACTION-FIELD ELECTROSTATICS */
1696             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
1697
1698             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
1699
1700             fscal            = felec;
1701
1702             fscal            = _mm_and_pd(fscal,cutoff_mask);
1703
1704             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1705
1706             /* Update vectorial force */
1707             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1708             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1709             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1710
1711             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1712             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1713             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1714
1715             }
1716
1717             /**************************
1718              * CALCULATE INTERACTIONS *
1719              **************************/
1720
1721             if (gmx_mm_any_lt(rsq11,rcutoff2))
1722             {
1723
1724             /* REACTION-FIELD ELECTROSTATICS */
1725             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1726
1727             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1728
1729             fscal            = felec;
1730
1731             fscal            = _mm_and_pd(fscal,cutoff_mask);
1732
1733             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1734
1735             /* Update vectorial force */
1736             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1737             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1738             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1739
1740             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1741             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1742             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1743
1744             }
1745
1746             /**************************
1747              * CALCULATE INTERACTIONS *
1748              **************************/
1749
1750             if (gmx_mm_any_lt(rsq12,rcutoff2))
1751             {
1752
1753             /* REACTION-FIELD ELECTROSTATICS */
1754             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1755
1756             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1757
1758             fscal            = felec;
1759
1760             fscal            = _mm_and_pd(fscal,cutoff_mask);
1761
1762             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1763
1764             /* Update vectorial force */
1765             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1766             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1767             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1768
1769             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1770             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1771             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1772
1773             }
1774
1775             /**************************
1776              * CALCULATE INTERACTIONS *
1777              **************************/
1778
1779             if (gmx_mm_any_lt(rsq20,rcutoff2))
1780             {
1781
1782             /* REACTION-FIELD ELECTROSTATICS */
1783             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
1784
1785             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
1786
1787             fscal            = felec;
1788
1789             fscal            = _mm_and_pd(fscal,cutoff_mask);
1790
1791             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1792
1793             /* Update vectorial force */
1794             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1795             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1796             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1797
1798             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1799             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1800             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1801
1802             }
1803
1804             /**************************
1805              * CALCULATE INTERACTIONS *
1806              **************************/
1807
1808             if (gmx_mm_any_lt(rsq21,rcutoff2))
1809             {
1810
1811             /* REACTION-FIELD ELECTROSTATICS */
1812             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1813
1814             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1815
1816             fscal            = felec;
1817
1818             fscal            = _mm_and_pd(fscal,cutoff_mask);
1819
1820             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1821
1822             /* Update vectorial force */
1823             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1824             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1825             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1826
1827             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1828             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1829             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1830
1831             }
1832
1833             /**************************
1834              * CALCULATE INTERACTIONS *
1835              **************************/
1836
1837             if (gmx_mm_any_lt(rsq22,rcutoff2))
1838             {
1839
1840             /* REACTION-FIELD ELECTROSTATICS */
1841             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1842
1843             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1844
1845             fscal            = felec;
1846
1847             fscal            = _mm_and_pd(fscal,cutoff_mask);
1848
1849             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1850
1851             /* Update vectorial force */
1852             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1853             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1854             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1855
1856             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1857             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1858             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1859
1860             }
1861
1862             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1863
1864             /* Inner loop uses 304 flops */
1865         }
1866
1867         /* End of innermost loop */
1868
1869         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1870                                               f+i_coord_offset,fshift+i_shift_offset);
1871
1872         /* Increment number of inner iterations */
1873         inneriter                  += j_index_end - j_index_start;
1874
1875         /* Outer loop uses 18 flops */
1876     }
1877
1878     /* Increment number of outer iterations */
1879     outeriter        += nri;
1880
1881     /* Update outer/inner flops */
1882
1883     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*304);
1884 }