src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecRFCut_VdwLJSw_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_VdwLJSw_GeomW4W4_VF_avx_128_fma_double
  54  * Electrostatics interaction: ReactionField
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Water4-Water4
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecRFCut_VdwLJSw_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              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              vdwioffset3;
  89     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  90     int              vdwjidx0A,vdwjidx0B;
  91     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  92     int              vdwjidx1A,vdwjidx1B;
  93     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  94     int              vdwjidx2A,vdwjidx2B;
  95     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  96     int              vdwjidx3A,vdwjidx3B;
  97     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  98     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  99     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 100     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 101     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 102     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 103     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 104     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 105     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 106     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 107     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 108     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 109     real             *charge;
 110     int              nvdwtype;
 111     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 112     int              *vdwtype;
 113     real             *vdwparam;
 114     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 115     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 116     __m128d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 117     real             rswitch_scalar,d_scalar;
 118     __m128d          dummy_mask,cutoff_mask;
 119     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 120     __m128d          one     = _mm_set1_pd(1.0);
 121     __m128d          two     = _mm_set1_pd(2.0);
 122     x                = xx[0];
 123     f                = ff[0];
 124
 125     nri              = nlist->nri;
 126     iinr             = nlist->iinr;
 127     jindex           = nlist->jindex;
 128     jjnr             = nlist->jjnr;
 129     shiftidx         = nlist->shift;
 130     gid              = nlist->gid;
 131     shiftvec         = fr->shift_vec[0];
 132     fshift           = fr->fshift[0];
 133     facel            = _mm_set1_pd(fr->epsfac);
 134     charge           = mdatoms->chargeA;
 135     krf              = _mm_set1_pd(fr->ic->k_rf);
 136     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 137     crf              = _mm_set1_pd(fr->ic->c_rf);
 138     nvdwtype         = fr->ntype;
 139     vdwparam         = fr->nbfp;
 140     vdwtype          = mdatoms->typeA;
 141
 142     /* Setup water-specific parameters */
 143     inr              = nlist->iinr[0];
 144     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 145     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 146     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 147     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 148
 149     jq1              = _mm_set1_pd(charge[inr+1]);
 150     jq2              = _mm_set1_pd(charge[inr+2]);
 151     jq3              = _mm_set1_pd(charge[inr+3]);
 152     vdwjidx0A        = 2*vdwtype[inr+0];
 153     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 154     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 155     qq11             = _mm_mul_pd(iq1,jq1);
 156     qq12             = _mm_mul_pd(iq1,jq2);
 157     qq13             = _mm_mul_pd(iq1,jq3);
 158     qq21             = _mm_mul_pd(iq2,jq1);
 159     qq22             = _mm_mul_pd(iq2,jq2);
 160     qq23             = _mm_mul_pd(iq2,jq3);
 161     qq31             = _mm_mul_pd(iq3,jq1);
 162     qq32             = _mm_mul_pd(iq3,jq2);
 163     qq33             = _mm_mul_pd(iq3,jq3);
 164
 165     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 166     rcutoff_scalar   = fr->rcoulomb;
 167     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 168     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 169
 170     rswitch_scalar   = fr->rvdw_switch;
 171     rswitch          = _mm_set1_pd(rswitch_scalar);
 172     /* Setup switch parameters */
 173     d_scalar         = rcutoff_scalar-rswitch_scalar;
 174     d                = _mm_set1_pd(d_scalar);
 175     swV3             = _mm_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 176     swV4             = _mm_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 177     swV5             = _mm_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 178     swF2             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 179     swF3             = _mm_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 180     swF4             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 181
 182     /* Avoid stupid compiler warnings */
 183     jnrA = jnrB = 0;
 184     j_coord_offsetA = 0;
 185     j_coord_offsetB = 0;
 186
 187     outeriter        = 0;
 188     inneriter        = 0;
 189
 190     /* Start outer loop over neighborlists */
 191     for(iidx=0; iidx<nri; iidx++)
 192     {
 193         /* Load shift vector for this list */
 194         i_shift_offset   = DIM*shiftidx[iidx];
 195
 196         /* Load limits for loop over neighbors */
 197         j_index_start    = jindex[iidx];
 198         j_index_end      = jindex[iidx+1];
 199
 200         /* Get outer coordinate index */
 201         inr              = iinr[iidx];
 202         i_coord_offset   = DIM*inr;
 203
 204         /* Load i particle coords and add shift vector */
 205         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 206                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 207
 208         fix0             = _mm_setzero_pd();
 209         fiy0             = _mm_setzero_pd();
 210         fiz0             = _mm_setzero_pd();
 211         fix1             = _mm_setzero_pd();
 212         fiy1             = _mm_setzero_pd();
 213         fiz1             = _mm_setzero_pd();
 214         fix2             = _mm_setzero_pd();
 215         fiy2             = _mm_setzero_pd();
 216         fiz2             = _mm_setzero_pd();
 217         fix3             = _mm_setzero_pd();
 218         fiy3             = _mm_setzero_pd();
 219         fiz3             = _mm_setzero_pd();
 220
 221         /* Reset potential sums */
 222         velecsum         = _mm_setzero_pd();
 223         vvdwsum          = _mm_setzero_pd();
 224
 225         /* Start inner kernel loop */
 226         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 227         {
 228
 229             /* Get j neighbor index, and coordinate index */
 230             jnrA             = jjnr[jidx];
 231             jnrB             = jjnr[jidx+1];
 232             j_coord_offsetA  = DIM*jnrA;
 233             j_coord_offsetB  = DIM*jnrB;
 234
 235             /* load j atom coordinates */
 236             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 237                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 238                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 239
 240             /* Calculate displacement vector */
 241             dx00             = _mm_sub_pd(ix0,jx0);
 242             dy00             = _mm_sub_pd(iy0,jy0);
 243             dz00             = _mm_sub_pd(iz0,jz0);
 244             dx11             = _mm_sub_pd(ix1,jx1);
 245             dy11             = _mm_sub_pd(iy1,jy1);
 246             dz11             = _mm_sub_pd(iz1,jz1);
 247             dx12             = _mm_sub_pd(ix1,jx2);
 248             dy12             = _mm_sub_pd(iy1,jy2);
 249             dz12             = _mm_sub_pd(iz1,jz2);
 250             dx13             = _mm_sub_pd(ix1,jx3);
 251             dy13             = _mm_sub_pd(iy1,jy3);
 252             dz13             = _mm_sub_pd(iz1,jz3);
 253             dx21             = _mm_sub_pd(ix2,jx1);
 254             dy21             = _mm_sub_pd(iy2,jy1);
 255             dz21             = _mm_sub_pd(iz2,jz1);
 256             dx22             = _mm_sub_pd(ix2,jx2);
 257             dy22             = _mm_sub_pd(iy2,jy2);
 258             dz22             = _mm_sub_pd(iz2,jz2);
 259             dx23             = _mm_sub_pd(ix2,jx3);
 260             dy23             = _mm_sub_pd(iy2,jy3);
 261             dz23             = _mm_sub_pd(iz2,jz3);
 262             dx31             = _mm_sub_pd(ix3,jx1);
 263             dy31             = _mm_sub_pd(iy3,jy1);
 264             dz31             = _mm_sub_pd(iz3,jz1);
 265             dx32             = _mm_sub_pd(ix3,jx2);
 266             dy32             = _mm_sub_pd(iy3,jy2);
 267             dz32             = _mm_sub_pd(iz3,jz2);
 268             dx33             = _mm_sub_pd(ix3,jx3);
 269             dy33             = _mm_sub_pd(iy3,jy3);
 270             dz33             = _mm_sub_pd(iz3,jz3);
 271
 272             /* Calculate squared distance and things based on it */
 273             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 274             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 275             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 276             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 277             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 278             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 279             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 280             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 281             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 282             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 283
 284             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 285             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 286             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 287             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 288             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 289             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 290             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 291             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 292             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 293             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 294
 295             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 296             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 297             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 298             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 299             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 300             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 301             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 302             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 303             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 304             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 305
 306             fjx0             = _mm_setzero_pd();
 307             fjy0             = _mm_setzero_pd();
 308             fjz0             = _mm_setzero_pd();
 309             fjx1             = _mm_setzero_pd();
 310             fjy1             = _mm_setzero_pd();
 311             fjz1             = _mm_setzero_pd();
 312             fjx2             = _mm_setzero_pd();
 313             fjy2             = _mm_setzero_pd();
 314             fjz2             = _mm_setzero_pd();
 315             fjx3             = _mm_setzero_pd();
 316             fjy3             = _mm_setzero_pd();
 317             fjz3             = _mm_setzero_pd();
 318
 319             /**************************
 320              * CALCULATE INTERACTIONS *
 321              **************************/
 322
 323             if (gmx_mm_any_lt(rsq00,rcutoff2))
 324             {
 325
 326             r00              = _mm_mul_pd(rsq00,rinv00);
 327
 328             /* LENNARD-JONES DISPERSION/REPULSION */
 329
 330             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 331             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 332             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 333             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 334             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 335
 336             d                = _mm_sub_pd(r00,rswitch);
 337             d                = _mm_max_pd(d,_mm_setzero_pd());
 338             d2               = _mm_mul_pd(d,d);
 339             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 340
 341             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 342
 343             /* Evaluate switch function */
 344             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 345             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 346             vvdw             = _mm_mul_pd(vvdw,sw);
 347             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 348
 349             /* Update potential sum for this i atom from the interaction with this j atom. */
 350             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 351             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 352
 353             fscal            = fvdw;
 354
 355             fscal            = _mm_and_pd(fscal,cutoff_mask);
 356
 357             /* Update vectorial force */
 358             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 359             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 360             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 361
 362             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 363             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 364             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 365
 366             }
 367
 368             /**************************
 369              * CALCULATE INTERACTIONS *
 370              **************************/
 371
 372             if (gmx_mm_any_lt(rsq11,rcutoff2))
 373             {
 374
 375             /* REACTION-FIELD ELECTROSTATICS */
 376             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 377             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 378
 379             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 380
 381             /* Update potential sum for this i atom from the interaction with this j atom. */
 382             velec            = _mm_and_pd(velec,cutoff_mask);
 383             velecsum         = _mm_add_pd(velecsum,velec);
 384
 385             fscal            = felec;
 386
 387             fscal            = _mm_and_pd(fscal,cutoff_mask);
 388
 389             /* Update vectorial force */
 390             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 391             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 392             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 393
 394             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 395             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 396             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 397
 398             }
 399
 400             /**************************
 401              * CALCULATE INTERACTIONS *
 402              **************************/
 403
 404             if (gmx_mm_any_lt(rsq12,rcutoff2))
 405             {
 406
 407             /* REACTION-FIELD ELECTROSTATICS */
 408             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 409             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 410
 411             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 412
 413             /* Update potential sum for this i atom from the interaction with this j atom. */
 414             velec            = _mm_and_pd(velec,cutoff_mask);
 415             velecsum         = _mm_add_pd(velecsum,velec);
 416
 417             fscal            = felec;
 418
 419             fscal            = _mm_and_pd(fscal,cutoff_mask);
 420
 421             /* Update vectorial force */
 422             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 423             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 424             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 425
 426             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 427             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 428             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 429
 430             }
 431
 432             /**************************
 433              * CALCULATE INTERACTIONS *
 434              **************************/
 435
 436             if (gmx_mm_any_lt(rsq13,rcutoff2))
 437             {
 438
 439             /* REACTION-FIELD ELECTROSTATICS */
 440             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 441             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 442
 443             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 444
 445             /* Update potential sum for this i atom from the interaction with this j atom. */
 446             velec            = _mm_and_pd(velec,cutoff_mask);
 447             velecsum         = _mm_add_pd(velecsum,velec);
 448
 449             fscal            = felec;
 450
 451             fscal            = _mm_and_pd(fscal,cutoff_mask);
 452
 453             /* Update vectorial force */
 454             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 455             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 456             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 457
 458             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 459             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 460             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 461
 462             }
 463
 464             /**************************
 465              * CALCULATE INTERACTIONS *
 466              **************************/
 467
 468             if (gmx_mm_any_lt(rsq21,rcutoff2))
 469             {
 470
 471             /* REACTION-FIELD ELECTROSTATICS */
 472             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 473             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 474
 475             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 476
 477             /* Update potential sum for this i atom from the interaction with this j atom. */
 478             velec            = _mm_and_pd(velec,cutoff_mask);
 479             velecsum         = _mm_add_pd(velecsum,velec);
 480
 481             fscal            = felec;
 482
 483             fscal            = _mm_and_pd(fscal,cutoff_mask);
 484
 485             /* Update vectorial force */
 486             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 487             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 488             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 489
 490             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 491             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 492             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 493
 494             }
 495
 496             /**************************
 497              * CALCULATE INTERACTIONS *
 498              **************************/
 499
 500             if (gmx_mm_any_lt(rsq22,rcutoff2))
 501             {
 502
 503             /* REACTION-FIELD ELECTROSTATICS */
 504             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 505             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 506
 507             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 508
 509             /* Update potential sum for this i atom from the interaction with this j atom. */
 510             velec            = _mm_and_pd(velec,cutoff_mask);
 511             velecsum         = _mm_add_pd(velecsum,velec);
 512
 513             fscal            = felec;
 514
 515             fscal            = _mm_and_pd(fscal,cutoff_mask);
 516
 517             /* Update vectorial force */
 518             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 519             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 520             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 521
 522             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 523             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 524             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 525
 526             }
 527
 528             /**************************
 529              * CALCULATE INTERACTIONS *
 530              **************************/
 531
 532             if (gmx_mm_any_lt(rsq23,rcutoff2))
 533             {
 534
 535             /* REACTION-FIELD ELECTROSTATICS */
 536             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 537             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 538
 539             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 540
 541             /* Update potential sum for this i atom from the interaction with this j atom. */
 542             velec            = _mm_and_pd(velec,cutoff_mask);
 543             velecsum         = _mm_add_pd(velecsum,velec);
 544
 545             fscal            = felec;
 546
 547             fscal            = _mm_and_pd(fscal,cutoff_mask);
 548
 549             /* Update vectorial force */
 550             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 551             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 552             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 553
 554             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 555             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 556             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 557
 558             }
 559
 560             /**************************
 561              * CALCULATE INTERACTIONS *
 562              **************************/
 563
 564             if (gmx_mm_any_lt(rsq31,rcutoff2))
 565             {
 566
 567             /* REACTION-FIELD ELECTROSTATICS */
 568             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
 569             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
 570
 571             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
 572
 573             /* Update potential sum for this i atom from the interaction with this j atom. */
 574             velec            = _mm_and_pd(velec,cutoff_mask);
 575             velecsum         = _mm_add_pd(velecsum,velec);
 576
 577             fscal            = felec;
 578
 579             fscal            = _mm_and_pd(fscal,cutoff_mask);
 580
 581             /* Update vectorial force */
 582             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 583             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 584             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 585
 586             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 587             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 588             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 589
 590             }
 591
 592             /**************************
 593              * CALCULATE INTERACTIONS *
 594              **************************/
 595
 596             if (gmx_mm_any_lt(rsq32,rcutoff2))
 597             {
 598
 599             /* REACTION-FIELD ELECTROSTATICS */
 600             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
 601             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
 602
 603             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
 604
 605             /* Update potential sum for this i atom from the interaction with this j atom. */
 606             velec            = _mm_and_pd(velec,cutoff_mask);
 607             velecsum         = _mm_add_pd(velecsum,velec);
 608
 609             fscal            = felec;
 610
 611             fscal            = _mm_and_pd(fscal,cutoff_mask);
 612
 613             /* Update vectorial force */
 614             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 615             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 616             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 617
 618             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 619             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 620             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 621
 622             }
 623
 624             /**************************
 625              * CALCULATE INTERACTIONS *
 626              **************************/
 627
 628             if (gmx_mm_any_lt(rsq33,rcutoff2))
 629             {
 630
 631             /* REACTION-FIELD ELECTROSTATICS */
 632             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
 633             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
 634
 635             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
 636
 637             /* Update potential sum for this i atom from the interaction with this j atom. */
 638             velec            = _mm_and_pd(velec,cutoff_mask);
 639             velecsum         = _mm_add_pd(velecsum,velec);
 640
 641             fscal            = felec;
 642
 643             fscal            = _mm_and_pd(fscal,cutoff_mask);
 644
 645             /* Update vectorial force */
 646             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 647             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 648             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 649
 650             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 651             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 652             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 653
 654             }
 655
 656             gmx_mm_decrement_4rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 657
 658             /* Inner loop uses 416 flops */
 659         }
 660
 661         if(jidx<j_index_end)
 662         {
 663
 664             jnrA             = jjnr[jidx];
 665             j_coord_offsetA  = DIM*jnrA;
 666
 667             /* load j atom coordinates */
 668             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 669                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 670                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 671
 672             /* Calculate displacement vector */
 673             dx00             = _mm_sub_pd(ix0,jx0);
 674             dy00             = _mm_sub_pd(iy0,jy0);
 675             dz00             = _mm_sub_pd(iz0,jz0);
 676             dx11             = _mm_sub_pd(ix1,jx1);
 677             dy11             = _mm_sub_pd(iy1,jy1);
 678             dz11             = _mm_sub_pd(iz1,jz1);
 679             dx12             = _mm_sub_pd(ix1,jx2);
 680             dy12             = _mm_sub_pd(iy1,jy2);
 681             dz12             = _mm_sub_pd(iz1,jz2);
 682             dx13             = _mm_sub_pd(ix1,jx3);
 683             dy13             = _mm_sub_pd(iy1,jy3);
 684             dz13             = _mm_sub_pd(iz1,jz3);
 685             dx21             = _mm_sub_pd(ix2,jx1);
 686             dy21             = _mm_sub_pd(iy2,jy1);
 687             dz21             = _mm_sub_pd(iz2,jz1);
 688             dx22             = _mm_sub_pd(ix2,jx2);
 689             dy22             = _mm_sub_pd(iy2,jy2);
 690             dz22             = _mm_sub_pd(iz2,jz2);
 691             dx23             = _mm_sub_pd(ix2,jx3);
 692             dy23             = _mm_sub_pd(iy2,jy3);
 693             dz23             = _mm_sub_pd(iz2,jz3);
 694             dx31             = _mm_sub_pd(ix3,jx1);
 695             dy31             = _mm_sub_pd(iy3,jy1);
 696             dz31             = _mm_sub_pd(iz3,jz1);
 697             dx32             = _mm_sub_pd(ix3,jx2);
 698             dy32             = _mm_sub_pd(iy3,jy2);
 699             dz32             = _mm_sub_pd(iz3,jz2);
 700             dx33             = _mm_sub_pd(ix3,jx3);
 701             dy33             = _mm_sub_pd(iy3,jy3);
 702             dz33             = _mm_sub_pd(iz3,jz3);
 703
 704             /* Calculate squared distance and things based on it */
 705             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 706             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 707             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 708             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 709             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 710             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 711             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 712             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 713             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 714             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 715
 716             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 717             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 718             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 719             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 720             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 721             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 722             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 723             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 724             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 725             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 726
 727             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 728             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 729             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 730             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 731             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 732             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 733             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 734             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 735             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 736             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 737
 738             fjx0             = _mm_setzero_pd();
 739             fjy0             = _mm_setzero_pd();
 740             fjz0             = _mm_setzero_pd();
 741             fjx1             = _mm_setzero_pd();
 742             fjy1             = _mm_setzero_pd();
 743             fjz1             = _mm_setzero_pd();
 744             fjx2             = _mm_setzero_pd();
 745             fjy2             = _mm_setzero_pd();
 746             fjz2             = _mm_setzero_pd();
 747             fjx3             = _mm_setzero_pd();
 748             fjy3             = _mm_setzero_pd();
 749             fjz3             = _mm_setzero_pd();
 750
 751             /**************************
 752              * CALCULATE INTERACTIONS *
 753              **************************/
 754
 755             if (gmx_mm_any_lt(rsq00,rcutoff2))
 756             {
 757
 758             r00              = _mm_mul_pd(rsq00,rinv00);
 759
 760             /* LENNARD-JONES DISPERSION/REPULSION */
 761
 762             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 763             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 764             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 765             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 766             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 767
 768             d                = _mm_sub_pd(r00,rswitch);
 769             d                = _mm_max_pd(d,_mm_setzero_pd());
 770             d2               = _mm_mul_pd(d,d);
 771             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 772
 773             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 774
 775             /* Evaluate switch function */
 776             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 777             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 778             vvdw             = _mm_mul_pd(vvdw,sw);
 779             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 780
 781             /* Update potential sum for this i atom from the interaction with this j atom. */
 782             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 783             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 784             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 785
 786             fscal            = fvdw;
 787
 788             fscal            = _mm_and_pd(fscal,cutoff_mask);
 789
 790             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 791
 792             /* Update vectorial force */
 793             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 794             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 795             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 796
 797             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 798             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 799             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 800
 801             }
 802
 803             /**************************
 804              * CALCULATE INTERACTIONS *
 805              **************************/
 806
 807             if (gmx_mm_any_lt(rsq11,rcutoff2))
 808             {
 809
 810             /* REACTION-FIELD ELECTROSTATICS */
 811             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 812             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 813
 814             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 815
 816             /* Update potential sum for this i atom from the interaction with this j atom. */
 817             velec            = _mm_and_pd(velec,cutoff_mask);
 818             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 819             velecsum         = _mm_add_pd(velecsum,velec);
 820
 821             fscal            = felec;
 822
 823             fscal            = _mm_and_pd(fscal,cutoff_mask);
 824
 825             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 826
 827             /* Update vectorial force */
 828             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 829             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 830             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 831
 832             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 833             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 834             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 835
 836             }
 837
 838             /**************************
 839              * CALCULATE INTERACTIONS *
 840              **************************/
 841
 842             if (gmx_mm_any_lt(rsq12,rcutoff2))
 843             {
 844
 845             /* REACTION-FIELD ELECTROSTATICS */
 846             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 847             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 848
 849             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 850
 851             /* Update potential sum for this i atom from the interaction with this j atom. */
 852             velec            = _mm_and_pd(velec,cutoff_mask);
 853             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 854             velecsum         = _mm_add_pd(velecsum,velec);
 855
 856             fscal            = felec;
 857
 858             fscal            = _mm_and_pd(fscal,cutoff_mask);
 859
 860             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 861
 862             /* Update vectorial force */
 863             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 864             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 865             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 866
 867             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 868             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 869             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 870
 871             }
 872
 873             /**************************
 874              * CALCULATE INTERACTIONS *
 875              **************************/
 876
 877             if (gmx_mm_any_lt(rsq13,rcutoff2))
 878             {
 879
 880             /* REACTION-FIELD ELECTROSTATICS */
 881             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 882             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 883
 884             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 885
 886             /* Update potential sum for this i atom from the interaction with this j atom. */
 887             velec            = _mm_and_pd(velec,cutoff_mask);
 888             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 889             velecsum         = _mm_add_pd(velecsum,velec);
 890
 891             fscal            = felec;
 892
 893             fscal            = _mm_and_pd(fscal,cutoff_mask);
 894
 895             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 896
 897             /* Update vectorial force */
 898             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 899             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 900             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 901
 902             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 903             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 904             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 905
 906             }
 907
 908             /**************************
 909              * CALCULATE INTERACTIONS *
 910              **************************/
 911
 912             if (gmx_mm_any_lt(rsq21,rcutoff2))
 913             {
 914
 915             /* REACTION-FIELD ELECTROSTATICS */
 916             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 917             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 918
 919             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 920
 921             /* Update potential sum for this i atom from the interaction with this j atom. */
 922             velec            = _mm_and_pd(velec,cutoff_mask);
 923             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 924             velecsum         = _mm_add_pd(velecsum,velec);
 925
 926             fscal            = felec;
 927
 928             fscal            = _mm_and_pd(fscal,cutoff_mask);
 929
 930             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 931
 932             /* Update vectorial force */
 933             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 934             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 935             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 936
 937             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 938             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 939             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 940
 941             }
 942
 943             /**************************
 944              * CALCULATE INTERACTIONS *
 945              **************************/
 946
 947             if (gmx_mm_any_lt(rsq22,rcutoff2))
 948             {
 949
 950             /* REACTION-FIELD ELECTROSTATICS */
 951             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 952             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 953
 954             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 955
 956             /* Update potential sum for this i atom from the interaction with this j atom. */
 957             velec            = _mm_and_pd(velec,cutoff_mask);
 958             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 959             velecsum         = _mm_add_pd(velecsum,velec);
 960
 961             fscal            = felec;
 962
 963             fscal            = _mm_and_pd(fscal,cutoff_mask);
 964
 965             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 966
 967             /* Update vectorial force */
 968             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 969             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 970             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 971
 972             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 973             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 974             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 975
 976             }
 977
 978             /**************************
 979              * CALCULATE INTERACTIONS *
 980              **************************/
 981
 982             if (gmx_mm_any_lt(rsq23,rcutoff2))
 983             {
 984
 985             /* REACTION-FIELD ELECTROSTATICS */
 986             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 987             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 988
 989             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 990
 991             /* Update potential sum for this i atom from the interaction with this j atom. */
 992             velec            = _mm_and_pd(velec,cutoff_mask);
 993             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 994             velecsum         = _mm_add_pd(velecsum,velec);
 995
 996             fscal            = felec;
 997
 998             fscal            = _mm_and_pd(fscal,cutoff_mask);
 999
1000             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1001
1002             /* Update vectorial force */
1003             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1004             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1005             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1006
1007             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1008             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1009             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1010
1011             }
1012
1013             /**************************
1014              * CALCULATE INTERACTIONS *
1015              **************************/
1016
1017             if (gmx_mm_any_lt(rsq31,rcutoff2))
1018             {
1019
1020             /* REACTION-FIELD ELECTROSTATICS */
1021             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
1022             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1023
1024             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1025
1026             /* Update potential sum for this i atom from the interaction with this j atom. */
1027             velec            = _mm_and_pd(velec,cutoff_mask);
1028             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1029             velecsum         = _mm_add_pd(velecsum,velec);
1030
1031             fscal            = felec;
1032
1033             fscal            = _mm_and_pd(fscal,cutoff_mask);
1034
1035             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1036
1037             /* Update vectorial force */
1038             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1039             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1040             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1041
1042             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1043             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1044             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1045
1046             }
1047
1048             /**************************
1049              * CALCULATE INTERACTIONS *
1050              **************************/
1051
1052             if (gmx_mm_any_lt(rsq32,rcutoff2))
1053             {
1054
1055             /* REACTION-FIELD ELECTROSTATICS */
1056             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
1057             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1058
1059             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1060
1061             /* Update potential sum for this i atom from the interaction with this j atom. */
1062             velec            = _mm_and_pd(velec,cutoff_mask);
1063             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1064             velecsum         = _mm_add_pd(velecsum,velec);
1065
1066             fscal            = felec;
1067
1068             fscal            = _mm_and_pd(fscal,cutoff_mask);
1069
1070             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1071
1072             /* Update vectorial force */
1073             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1074             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1075             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1076
1077             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1078             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1079             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1080
1081             }
1082
1083             /**************************
1084              * CALCULATE INTERACTIONS *
1085              **************************/
1086
1087             if (gmx_mm_any_lt(rsq33,rcutoff2))
1088             {
1089
1090             /* REACTION-FIELD ELECTROSTATICS */
1091             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
1092             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1093
1094             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1095
1096             /* Update potential sum for this i atom from the interaction with this j atom. */
1097             velec            = _mm_and_pd(velec,cutoff_mask);
1098             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1099             velecsum         = _mm_add_pd(velecsum,velec);
1100
1101             fscal            = felec;
1102
1103             fscal            = _mm_and_pd(fscal,cutoff_mask);
1104
1105             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1106
1107             /* Update vectorial force */
1108             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1109             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1110             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1111
1112             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1113             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1114             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1115
1116             }
1117
1118             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1119
1120             /* Inner loop uses 416 flops */
1121         }
1122
1123         /* End of innermost loop */
1124
1125         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1126                                               f+i_coord_offset,fshift+i_shift_offset);
1127
1128         ggid                        = gid[iidx];
1129         /* Update potential energies */
1130         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1131         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
1132
1133         /* Increment number of inner iterations */
1134         inneriter                  += j_index_end - j_index_start;
1135
1136         /* Outer loop uses 26 flops */
1137     }
1138
1139     /* Increment number of outer iterations */
1140     outeriter        += nri;
1141
1142     /* Update outer/inner flops */
1143
1144     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*416);
1145 }
1146 /*
1147  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_F_avx_128_fma_double
1148  * Electrostatics interaction: ReactionField
1149  * VdW interaction:            LennardJones
1150  * Geometry:                   Water4-Water4
1151  * Calculate force/pot:        Force
1152  */
1153 void
1154 nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_F_avx_128_fma_double
1155                     (t_nblist                    * gmx_restrict       nlist,
1156                      rvec                        * gmx_restrict          xx,
1157                      rvec                        * gmx_restrict          ff,
1158                      t_forcerec                  * gmx_restrict          fr,
1159                      t_mdatoms                   * gmx_restrict     mdatoms,
1160                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1161                      t_nrnb                      * gmx_restrict        nrnb)
1162 {
1163     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1164      * just 0 for non-waters.
1165      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1166      * jnr indices corresponding to data put in the four positions in the SIMD register.
1167      */
1168     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1169     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1170     int              jnrA,jnrB;
1171     int              j_coord_offsetA,j_coord_offsetB;
1172     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1173     real             rcutoff_scalar;
1174     real             *shiftvec,*fshift,*x,*f;
1175     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1176     int              vdwioffset0;
1177     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1178     int              vdwioffset1;
1179     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1180     int              vdwioffset2;
1181     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1182     int              vdwioffset3;
1183     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1184     int              vdwjidx0A,vdwjidx0B;
1185     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1186     int              vdwjidx1A,vdwjidx1B;
1187     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1188     int              vdwjidx2A,vdwjidx2B;
1189     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1190     int              vdwjidx3A,vdwjidx3B;
1191     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1192     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1193     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1194     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1195     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1196     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1197     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1198     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1199     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1200     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1201     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1202     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1203     real             *charge;
1204     int              nvdwtype;
1205     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1206     int              *vdwtype;
1207     real             *vdwparam;
1208     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1209     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1210     __m128d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
1211     real             rswitch_scalar,d_scalar;
1212     __m128d          dummy_mask,cutoff_mask;
1213     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1214     __m128d          one     = _mm_set1_pd(1.0);
1215     __m128d          two     = _mm_set1_pd(2.0);
1216     x                = xx[0];
1217     f                = ff[0];
1218
1219     nri              = nlist->nri;
1220     iinr             = nlist->iinr;
1221     jindex           = nlist->jindex;
1222     jjnr             = nlist->jjnr;
1223     shiftidx         = nlist->shift;
1224     gid              = nlist->gid;
1225     shiftvec         = fr->shift_vec[0];
1226     fshift           = fr->fshift[0];
1227     facel            = _mm_set1_pd(fr->epsfac);
1228     charge           = mdatoms->chargeA;
1229     krf              = _mm_set1_pd(fr->ic->k_rf);
1230     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1231     crf              = _mm_set1_pd(fr->ic->c_rf);
1232     nvdwtype         = fr->ntype;
1233     vdwparam         = fr->nbfp;
1234     vdwtype          = mdatoms->typeA;
1235
1236     /* Setup water-specific parameters */
1237     inr              = nlist->iinr[0];
1238     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1239     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1240     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1241     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1242
1243     jq1              = _mm_set1_pd(charge[inr+1]);
1244     jq2              = _mm_set1_pd(charge[inr+2]);
1245     jq3              = _mm_set1_pd(charge[inr+3]);
1246     vdwjidx0A        = 2*vdwtype[inr+0];
1247     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1248     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1249     qq11             = _mm_mul_pd(iq1,jq1);
1250     qq12             = _mm_mul_pd(iq1,jq2);
1251     qq13             = _mm_mul_pd(iq1,jq3);
1252     qq21             = _mm_mul_pd(iq2,jq1);
1253     qq22             = _mm_mul_pd(iq2,jq2);
1254     qq23             = _mm_mul_pd(iq2,jq3);
1255     qq31             = _mm_mul_pd(iq3,jq1);
1256     qq32             = _mm_mul_pd(iq3,jq2);
1257     qq33             = _mm_mul_pd(iq3,jq3);
1258
1259     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1260     rcutoff_scalar   = fr->rcoulomb;
1261     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1262     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1263
1264     rswitch_scalar   = fr->rvdw_switch;
1265     rswitch          = _mm_set1_pd(rswitch_scalar);
1266     /* Setup switch parameters */
1267     d_scalar         = rcutoff_scalar-rswitch_scalar;
1268     d                = _mm_set1_pd(d_scalar);
1269     swV3             = _mm_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
1270     swV4             = _mm_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1271     swV5             = _mm_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1272     swF2             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
1273     swF3             = _mm_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1274     swF4             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1275
1276     /* Avoid stupid compiler warnings */
1277     jnrA = jnrB = 0;
1278     j_coord_offsetA = 0;
1279     j_coord_offsetB = 0;
1280
1281     outeriter        = 0;
1282     inneriter        = 0;
1283
1284     /* Start outer loop over neighborlists */
1285     for(iidx=0; iidx<nri; iidx++)
1286     {
1287         /* Load shift vector for this list */
1288         i_shift_offset   = DIM*shiftidx[iidx];
1289
1290         /* Load limits for loop over neighbors */
1291         j_index_start    = jindex[iidx];
1292         j_index_end      = jindex[iidx+1];
1293
1294         /* Get outer coordinate index */
1295         inr              = iinr[iidx];
1296         i_coord_offset   = DIM*inr;
1297
1298         /* Load i particle coords and add shift vector */
1299         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1300                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1301
1302         fix0             = _mm_setzero_pd();
1303         fiy0             = _mm_setzero_pd();
1304         fiz0             = _mm_setzero_pd();
1305         fix1             = _mm_setzero_pd();
1306         fiy1             = _mm_setzero_pd();
1307         fiz1             = _mm_setzero_pd();
1308         fix2             = _mm_setzero_pd();
1309         fiy2             = _mm_setzero_pd();
1310         fiz2             = _mm_setzero_pd();
1311         fix3             = _mm_setzero_pd();
1312         fiy3             = _mm_setzero_pd();
1313         fiz3             = _mm_setzero_pd();
1314
1315         /* Start inner kernel loop */
1316         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1317         {
1318
1319             /* Get j neighbor index, and coordinate index */
1320             jnrA             = jjnr[jidx];
1321             jnrB             = jjnr[jidx+1];
1322             j_coord_offsetA  = DIM*jnrA;
1323             j_coord_offsetB  = DIM*jnrB;
1324
1325             /* load j atom coordinates */
1326             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1327                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1328                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1329
1330             /* Calculate displacement vector */
1331             dx00             = _mm_sub_pd(ix0,jx0);
1332             dy00             = _mm_sub_pd(iy0,jy0);
1333             dz00             = _mm_sub_pd(iz0,jz0);
1334             dx11             = _mm_sub_pd(ix1,jx1);
1335             dy11             = _mm_sub_pd(iy1,jy1);
1336             dz11             = _mm_sub_pd(iz1,jz1);
1337             dx12             = _mm_sub_pd(ix1,jx2);
1338             dy12             = _mm_sub_pd(iy1,jy2);
1339             dz12             = _mm_sub_pd(iz1,jz2);
1340             dx13             = _mm_sub_pd(ix1,jx3);
1341             dy13             = _mm_sub_pd(iy1,jy3);
1342             dz13             = _mm_sub_pd(iz1,jz3);
1343             dx21             = _mm_sub_pd(ix2,jx1);
1344             dy21             = _mm_sub_pd(iy2,jy1);
1345             dz21             = _mm_sub_pd(iz2,jz1);
1346             dx22             = _mm_sub_pd(ix2,jx2);
1347             dy22             = _mm_sub_pd(iy2,jy2);
1348             dz22             = _mm_sub_pd(iz2,jz2);
1349             dx23             = _mm_sub_pd(ix2,jx3);
1350             dy23             = _mm_sub_pd(iy2,jy3);
1351             dz23             = _mm_sub_pd(iz2,jz3);
1352             dx31             = _mm_sub_pd(ix3,jx1);
1353             dy31             = _mm_sub_pd(iy3,jy1);
1354             dz31             = _mm_sub_pd(iz3,jz1);
1355             dx32             = _mm_sub_pd(ix3,jx2);
1356             dy32             = _mm_sub_pd(iy3,jy2);
1357             dz32             = _mm_sub_pd(iz3,jz2);
1358             dx33             = _mm_sub_pd(ix3,jx3);
1359             dy33             = _mm_sub_pd(iy3,jy3);
1360             dz33             = _mm_sub_pd(iz3,jz3);
1361
1362             /* Calculate squared distance and things based on it */
1363             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1364             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1365             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1366             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1367             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1368             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1369             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1370             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1371             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1372             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1373
1374             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1375             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1376             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1377             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1378             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1379             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1380             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1381             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1382             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1383             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1384
1385             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1386             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1387             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1388             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1389             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1390             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1391             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1392             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1393             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1394             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1395
1396             fjx0             = _mm_setzero_pd();
1397             fjy0             = _mm_setzero_pd();
1398             fjz0             = _mm_setzero_pd();
1399             fjx1             = _mm_setzero_pd();
1400             fjy1             = _mm_setzero_pd();
1401             fjz1             = _mm_setzero_pd();
1402             fjx2             = _mm_setzero_pd();
1403             fjy2             = _mm_setzero_pd();
1404             fjz2             = _mm_setzero_pd();
1405             fjx3             = _mm_setzero_pd();
1406             fjy3             = _mm_setzero_pd();
1407             fjz3             = _mm_setzero_pd();
1408
1409             /**************************
1410              * CALCULATE INTERACTIONS *
1411              **************************/
1412
1413             if (gmx_mm_any_lt(rsq00,rcutoff2))
1414             {
1415
1416             r00              = _mm_mul_pd(rsq00,rinv00);
1417
1418             /* LENNARD-JONES DISPERSION/REPULSION */
1419
1420             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1421             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
1422             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
1423             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
1424             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
1425
1426             d                = _mm_sub_pd(r00,rswitch);
1427             d                = _mm_max_pd(d,_mm_setzero_pd());
1428             d2               = _mm_mul_pd(d,d);
1429             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
1430
1431             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
1432
1433             /* Evaluate switch function */
1434             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1435             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
1436             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1437
1438             fscal            = fvdw;
1439
1440             fscal            = _mm_and_pd(fscal,cutoff_mask);
1441
1442             /* Update vectorial force */
1443             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1444             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1445             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1446
1447             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1448             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1449             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1450
1451             }
1452
1453             /**************************
1454              * CALCULATE INTERACTIONS *
1455              **************************/
1456
1457             if (gmx_mm_any_lt(rsq11,rcutoff2))
1458             {
1459
1460             /* REACTION-FIELD ELECTROSTATICS */
1461             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1462
1463             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1464
1465             fscal            = felec;
1466
1467             fscal            = _mm_and_pd(fscal,cutoff_mask);
1468
1469             /* Update vectorial force */
1470             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1471             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1472             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1473
1474             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1475             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1476             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1477
1478             }
1479
1480             /**************************
1481              * CALCULATE INTERACTIONS *
1482              **************************/
1483
1484             if (gmx_mm_any_lt(rsq12,rcutoff2))
1485             {
1486
1487             /* REACTION-FIELD ELECTROSTATICS */
1488             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1489
1490             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1491
1492             fscal            = felec;
1493
1494             fscal            = _mm_and_pd(fscal,cutoff_mask);
1495
1496             /* Update vectorial force */
1497             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1498             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1499             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1500
1501             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1502             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1503             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1504
1505             }
1506
1507             /**************************
1508              * CALCULATE INTERACTIONS *
1509              **************************/
1510
1511             if (gmx_mm_any_lt(rsq13,rcutoff2))
1512             {
1513
1514             /* REACTION-FIELD ELECTROSTATICS */
1515             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1516
1517             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1518
1519             fscal            = felec;
1520
1521             fscal            = _mm_and_pd(fscal,cutoff_mask);
1522
1523             /* Update vectorial force */
1524             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1525             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1526             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1527
1528             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1529             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1530             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1531
1532             }
1533
1534             /**************************
1535              * CALCULATE INTERACTIONS *
1536              **************************/
1537
1538             if (gmx_mm_any_lt(rsq21,rcutoff2))
1539             {
1540
1541             /* REACTION-FIELD ELECTROSTATICS */
1542             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1543
1544             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1545
1546             fscal            = felec;
1547
1548             fscal            = _mm_and_pd(fscal,cutoff_mask);
1549
1550             /* Update vectorial force */
1551             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1552             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1553             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1554
1555             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1556             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1557             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1558
1559             }
1560
1561             /**************************
1562              * CALCULATE INTERACTIONS *
1563              **************************/
1564
1565             if (gmx_mm_any_lt(rsq22,rcutoff2))
1566             {
1567
1568             /* REACTION-FIELD ELECTROSTATICS */
1569             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1570
1571             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1572
1573             fscal            = felec;
1574
1575             fscal            = _mm_and_pd(fscal,cutoff_mask);
1576
1577             /* Update vectorial force */
1578             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1579             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1580             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1581
1582             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1583             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1584             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1585
1586             }
1587
1588             /**************************
1589              * CALCULATE INTERACTIONS *
1590              **************************/
1591
1592             if (gmx_mm_any_lt(rsq23,rcutoff2))
1593             {
1594
1595             /* REACTION-FIELD ELECTROSTATICS */
1596             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1597
1598             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1599
1600             fscal            = felec;
1601
1602             fscal            = _mm_and_pd(fscal,cutoff_mask);
1603
1604             /* Update vectorial force */
1605             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1606             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1607             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1608
1609             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1610             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1611             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1612
1613             }
1614
1615             /**************************
1616              * CALCULATE INTERACTIONS *
1617              **************************/
1618
1619             if (gmx_mm_any_lt(rsq31,rcutoff2))
1620             {
1621
1622             /* REACTION-FIELD ELECTROSTATICS */
1623             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1624
1625             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1626
1627             fscal            = felec;
1628
1629             fscal            = _mm_and_pd(fscal,cutoff_mask);
1630
1631             /* Update vectorial force */
1632             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1633             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1634             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1635
1636             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1637             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1638             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1639
1640             }
1641
1642             /**************************
1643              * CALCULATE INTERACTIONS *
1644              **************************/
1645
1646             if (gmx_mm_any_lt(rsq32,rcutoff2))
1647             {
1648
1649             /* REACTION-FIELD ELECTROSTATICS */
1650             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1651
1652             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1653
1654             fscal            = felec;
1655
1656             fscal            = _mm_and_pd(fscal,cutoff_mask);
1657
1658             /* Update vectorial force */
1659             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1660             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1661             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1662
1663             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1664             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1665             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1666
1667             }
1668
1669             /**************************
1670              * CALCULATE INTERACTIONS *
1671              **************************/
1672
1673             if (gmx_mm_any_lt(rsq33,rcutoff2))
1674             {
1675
1676             /* REACTION-FIELD ELECTROSTATICS */
1677             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1678
1679             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1680
1681             fscal            = felec;
1682
1683             fscal            = _mm_and_pd(fscal,cutoff_mask);
1684
1685             /* Update vectorial force */
1686             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1687             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1688             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1689
1690             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1691             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1692             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1693
1694             }
1695
1696             gmx_mm_decrement_4rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1697
1698             /* Inner loop uses 359 flops */
1699         }
1700
1701         if(jidx<j_index_end)
1702         {
1703
1704             jnrA             = jjnr[jidx];
1705             j_coord_offsetA  = DIM*jnrA;
1706
1707             /* load j atom coordinates */
1708             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1709                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1710                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1711
1712             /* Calculate displacement vector */
1713             dx00             = _mm_sub_pd(ix0,jx0);
1714             dy00             = _mm_sub_pd(iy0,jy0);
1715             dz00             = _mm_sub_pd(iz0,jz0);
1716             dx11             = _mm_sub_pd(ix1,jx1);
1717             dy11             = _mm_sub_pd(iy1,jy1);
1718             dz11             = _mm_sub_pd(iz1,jz1);
1719             dx12             = _mm_sub_pd(ix1,jx2);
1720             dy12             = _mm_sub_pd(iy1,jy2);
1721             dz12             = _mm_sub_pd(iz1,jz2);
1722             dx13             = _mm_sub_pd(ix1,jx3);
1723             dy13             = _mm_sub_pd(iy1,jy3);
1724             dz13             = _mm_sub_pd(iz1,jz3);
1725             dx21             = _mm_sub_pd(ix2,jx1);
1726             dy21             = _mm_sub_pd(iy2,jy1);
1727             dz21             = _mm_sub_pd(iz2,jz1);
1728             dx22             = _mm_sub_pd(ix2,jx2);
1729             dy22             = _mm_sub_pd(iy2,jy2);
1730             dz22             = _mm_sub_pd(iz2,jz2);
1731             dx23             = _mm_sub_pd(ix2,jx3);
1732             dy23             = _mm_sub_pd(iy2,jy3);
1733             dz23             = _mm_sub_pd(iz2,jz3);
1734             dx31             = _mm_sub_pd(ix3,jx1);
1735             dy31             = _mm_sub_pd(iy3,jy1);
1736             dz31             = _mm_sub_pd(iz3,jz1);
1737             dx32             = _mm_sub_pd(ix3,jx2);
1738             dy32             = _mm_sub_pd(iy3,jy2);
1739             dz32             = _mm_sub_pd(iz3,jz2);
1740             dx33             = _mm_sub_pd(ix3,jx3);
1741             dy33             = _mm_sub_pd(iy3,jy3);
1742             dz33             = _mm_sub_pd(iz3,jz3);
1743
1744             /* Calculate squared distance and things based on it */
1745             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1746             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1747             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1748             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1749             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1750             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1751             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1752             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1753             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1754             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1755
1756             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1757             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1758             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1759             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1760             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1761             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1762             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1763             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1764             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1765             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1766
1767             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1768             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1769             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1770             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1771             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1772             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1773             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1774             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1775             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1776             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1777
1778             fjx0             = _mm_setzero_pd();
1779             fjy0             = _mm_setzero_pd();
1780             fjz0             = _mm_setzero_pd();
1781             fjx1             = _mm_setzero_pd();
1782             fjy1             = _mm_setzero_pd();
1783             fjz1             = _mm_setzero_pd();
1784             fjx2             = _mm_setzero_pd();
1785             fjy2             = _mm_setzero_pd();
1786             fjz2             = _mm_setzero_pd();
1787             fjx3             = _mm_setzero_pd();
1788             fjy3             = _mm_setzero_pd();
1789             fjz3             = _mm_setzero_pd();
1790
1791             /**************************
1792              * CALCULATE INTERACTIONS *
1793              **************************/
1794
1795             if (gmx_mm_any_lt(rsq00,rcutoff2))
1796             {
1797
1798             r00              = _mm_mul_pd(rsq00,rinv00);
1799
1800             /* LENNARD-JONES DISPERSION/REPULSION */
1801
1802             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1803             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
1804             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
1805             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
1806             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
1807
1808             d                = _mm_sub_pd(r00,rswitch);
1809             d                = _mm_max_pd(d,_mm_setzero_pd());
1810             d2               = _mm_mul_pd(d,d);
1811             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
1812
1813             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
1814
1815             /* Evaluate switch function */
1816             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1817             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
1818             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1819
1820             fscal            = fvdw;
1821
1822             fscal            = _mm_and_pd(fscal,cutoff_mask);
1823
1824             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1825
1826             /* Update vectorial force */
1827             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1828             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1829             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1830
1831             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1832             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1833             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1834
1835             }
1836
1837             /**************************
1838              * CALCULATE INTERACTIONS *
1839              **************************/
1840
1841             if (gmx_mm_any_lt(rsq11,rcutoff2))
1842             {
1843
1844             /* REACTION-FIELD ELECTROSTATICS */
1845             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1846
1847             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1848
1849             fscal            = felec;
1850
1851             fscal            = _mm_and_pd(fscal,cutoff_mask);
1852
1853             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1854
1855             /* Update vectorial force */
1856             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1857             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1858             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1859
1860             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1861             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1862             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1863
1864             }
1865
1866             /**************************
1867              * CALCULATE INTERACTIONS *
1868              **************************/
1869
1870             if (gmx_mm_any_lt(rsq12,rcutoff2))
1871             {
1872
1873             /* REACTION-FIELD ELECTROSTATICS */
1874             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1875
1876             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1877
1878             fscal            = felec;
1879
1880             fscal            = _mm_and_pd(fscal,cutoff_mask);
1881
1882             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1883
1884             /* Update vectorial force */
1885             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1886             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1887             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1888
1889             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1890             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1891             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1892
1893             }
1894
1895             /**************************
1896              * CALCULATE INTERACTIONS *
1897              **************************/
1898
1899             if (gmx_mm_any_lt(rsq13,rcutoff2))
1900             {
1901
1902             /* REACTION-FIELD ELECTROSTATICS */
1903             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1904
1905             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1906
1907             fscal            = felec;
1908
1909             fscal            = _mm_and_pd(fscal,cutoff_mask);
1910
1911             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1912
1913             /* Update vectorial force */
1914             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1915             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1916             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1917
1918             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1919             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1920             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1921
1922             }
1923
1924             /**************************
1925              * CALCULATE INTERACTIONS *
1926              **************************/
1927
1928             if (gmx_mm_any_lt(rsq21,rcutoff2))
1929             {
1930
1931             /* REACTION-FIELD ELECTROSTATICS */
1932             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1933
1934             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1935
1936             fscal            = felec;
1937
1938             fscal            = _mm_and_pd(fscal,cutoff_mask);
1939
1940             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1941
1942             /* Update vectorial force */
1943             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1944             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1945             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1946
1947             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1948             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1949             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1950
1951             }
1952
1953             /**************************
1954              * CALCULATE INTERACTIONS *
1955              **************************/
1956
1957             if (gmx_mm_any_lt(rsq22,rcutoff2))
1958             {
1959
1960             /* REACTION-FIELD ELECTROSTATICS */
1961             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1962
1963             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1964
1965             fscal            = felec;
1966
1967             fscal            = _mm_and_pd(fscal,cutoff_mask);
1968
1969             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1970
1971             /* Update vectorial force */
1972             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1973             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1974             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1975
1976             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1977             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1978             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1979
1980             }
1981
1982             /**************************
1983              * CALCULATE INTERACTIONS *
1984              **************************/
1985
1986             if (gmx_mm_any_lt(rsq23,rcutoff2))
1987             {
1988
1989             /* REACTION-FIELD ELECTROSTATICS */
1990             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1991
1992             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1993
1994             fscal            = felec;
1995
1996             fscal            = _mm_and_pd(fscal,cutoff_mask);
1997
1998             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1999
2000             /* Update vectorial force */
2001             fix2             = _mm_macc_pd(dx23,fscal,fix2);
2002             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
2003             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
2004
2005             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
2006             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
2007             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
2008
2009             }
2010
2011             /**************************
2012              * CALCULATE INTERACTIONS *
2013              **************************/
2014
2015             if (gmx_mm_any_lt(rsq31,rcutoff2))
2016             {
2017
2018             /* REACTION-FIELD ELECTROSTATICS */
2019             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
2020
2021             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
2022
2023             fscal            = felec;
2024
2025             fscal            = _mm_and_pd(fscal,cutoff_mask);
2026
2027             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2028
2029             /* Update vectorial force */
2030             fix3             = _mm_macc_pd(dx31,fscal,fix3);
2031             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
2032             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
2033
2034             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
2035             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
2036             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
2037
2038             }
2039
2040             /**************************
2041              * CALCULATE INTERACTIONS *
2042              **************************/
2043
2044             if (gmx_mm_any_lt(rsq32,rcutoff2))
2045             {
2046
2047             /* REACTION-FIELD ELECTROSTATICS */
2048             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
2049
2050             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
2051
2052             fscal            = felec;
2053
2054             fscal            = _mm_and_pd(fscal,cutoff_mask);
2055
2056             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2057
2058             /* Update vectorial force */
2059             fix3             = _mm_macc_pd(dx32,fscal,fix3);
2060             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
2061             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
2062
2063             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
2064             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
2065             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
2066
2067             }
2068
2069             /**************************
2070              * CALCULATE INTERACTIONS *
2071              **************************/
2072
2073             if (gmx_mm_any_lt(rsq33,rcutoff2))
2074             {
2075
2076             /* REACTION-FIELD ELECTROSTATICS */
2077             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
2078
2079             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
2080
2081             fscal            = felec;
2082
2083             fscal            = _mm_and_pd(fscal,cutoff_mask);
2084
2085             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2086
2087             /* Update vectorial force */
2088             fix3             = _mm_macc_pd(dx33,fscal,fix3);
2089             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
2090             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
2091
2092             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
2093             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
2094             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
2095
2096             }
2097
2098             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2099
2100             /* Inner loop uses 359 flops */
2101         }
2102
2103         /* End of innermost loop */
2104
2105         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2106                                               f+i_coord_offset,fshift+i_shift_offset);
2107
2108         /* Increment number of inner iterations */
2109         inneriter                  += j_index_end - j_index_start;
2110
2111         /* Outer loop uses 24 flops */
2112     }
2113
2114     /* Increment number of outer iterations */
2115     outeriter        += nri;
2116
2117     /* Update outer/inner flops */
2118
2119     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*359);
2120 }