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) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_double kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_double.h"
  48 #include "kernelutil_x86_avx_128_fma_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_VF_avx_128_fma_double
  52  * Electrostatics interaction: ReactionField
  53  * VdW interaction:            LennardJones
  54  * Geometry:                   Water4-Water4
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_VF_avx_128_fma_double
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB;
  75     int              j_coord_offsetA,j_coord_offsetB;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  80     int              vdwioffset0;
  81     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  82     int              vdwioffset1;
  83     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  84     int              vdwioffset2;
  85     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  86     int              vdwioffset3;
  87     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  88     int              vdwjidx0A,vdwjidx0B;
  89     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     int              vdwjidx1A,vdwjidx1B;
  91     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  92     int              vdwjidx2A,vdwjidx2B;
  93     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  94     int              vdwjidx3A,vdwjidx3B;
  95     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  96     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  97     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  98     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  99     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 100     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 101     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 102     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 103     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 104     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 105     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 106     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 107     real             *charge;
 108     int              nvdwtype;
 109     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 110     int              *vdwtype;
 111     real             *vdwparam;
 112     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 113     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 114     __m128d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 115     real             rswitch_scalar,d_scalar;
 116     __m128d          dummy_mask,cutoff_mask;
 117     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 118     __m128d          one     = _mm_set1_pd(1.0);
 119     __m128d          two     = _mm_set1_pd(2.0);
 120     x                = xx[0];
 121     f                = ff[0];
 122
 123     nri              = nlist->nri;
 124     iinr             = nlist->iinr;
 125     jindex           = nlist->jindex;
 126     jjnr             = nlist->jjnr;
 127     shiftidx         = nlist->shift;
 128     gid              = nlist->gid;
 129     shiftvec         = fr->shift_vec[0];
 130     fshift           = fr->fshift[0];
 131     facel            = _mm_set1_pd(fr->epsfac);
 132     charge           = mdatoms->chargeA;
 133     krf              = _mm_set1_pd(fr->ic->k_rf);
 134     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 135     crf              = _mm_set1_pd(fr->ic->c_rf);
 136     nvdwtype         = fr->ntype;
 137     vdwparam         = fr->nbfp;
 138     vdwtype          = mdatoms->typeA;
 139
 140     /* Setup water-specific parameters */
 141     inr              = nlist->iinr[0];
 142     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 143     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 144     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 145     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 146
 147     jq1              = _mm_set1_pd(charge[inr+1]);
 148     jq2              = _mm_set1_pd(charge[inr+2]);
 149     jq3              = _mm_set1_pd(charge[inr+3]);
 150     vdwjidx0A        = 2*vdwtype[inr+0];
 151     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 152     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 153     qq11             = _mm_mul_pd(iq1,jq1);
 154     qq12             = _mm_mul_pd(iq1,jq2);
 155     qq13             = _mm_mul_pd(iq1,jq3);
 156     qq21             = _mm_mul_pd(iq2,jq1);
 157     qq22             = _mm_mul_pd(iq2,jq2);
 158     qq23             = _mm_mul_pd(iq2,jq3);
 159     qq31             = _mm_mul_pd(iq3,jq1);
 160     qq32             = _mm_mul_pd(iq3,jq2);
 161     qq33             = _mm_mul_pd(iq3,jq3);
 162
 163     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 164     rcutoff_scalar   = fr->rcoulomb;
 165     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 166     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 167
 168     rswitch_scalar   = fr->rvdw_switch;
 169     rswitch          = _mm_set1_pd(rswitch_scalar);
 170     /* Setup switch parameters */
 171     d_scalar         = rcutoff_scalar-rswitch_scalar;
 172     d                = _mm_set1_pd(d_scalar);
 173     swV3             = _mm_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 174     swV4             = _mm_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 175     swV5             = _mm_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 176     swF2             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 177     swF3             = _mm_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 178     swF4             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 179
 180     /* Avoid stupid compiler warnings */
 181     jnrA = jnrB = 0;
 182     j_coord_offsetA = 0;
 183     j_coord_offsetB = 0;
 184
 185     outeriter        = 0;
 186     inneriter        = 0;
 187
 188     /* Start outer loop over neighborlists */
 189     for(iidx=0; iidx<nri; iidx++)
 190     {
 191         /* Load shift vector for this list */
 192         i_shift_offset   = DIM*shiftidx[iidx];
 193
 194         /* Load limits for loop over neighbors */
 195         j_index_start    = jindex[iidx];
 196         j_index_end      = jindex[iidx+1];
 197
 198         /* Get outer coordinate index */
 199         inr              = iinr[iidx];
 200         i_coord_offset   = DIM*inr;
 201
 202         /* Load i particle coords and add shift vector */
 203         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 204                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 205
 206         fix0             = _mm_setzero_pd();
 207         fiy0             = _mm_setzero_pd();
 208         fiz0             = _mm_setzero_pd();
 209         fix1             = _mm_setzero_pd();
 210         fiy1             = _mm_setzero_pd();
 211         fiz1             = _mm_setzero_pd();
 212         fix2             = _mm_setzero_pd();
 213         fiy2             = _mm_setzero_pd();
 214         fiz2             = _mm_setzero_pd();
 215         fix3             = _mm_setzero_pd();
 216         fiy3             = _mm_setzero_pd();
 217         fiz3             = _mm_setzero_pd();
 218
 219         /* Reset potential sums */
 220         velecsum         = _mm_setzero_pd();
 221         vvdwsum          = _mm_setzero_pd();
 222
 223         /* Start inner kernel loop */
 224         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 225         {
 226
 227             /* Get j neighbor index, and coordinate index */
 228             jnrA             = jjnr[jidx];
 229             jnrB             = jjnr[jidx+1];
 230             j_coord_offsetA  = DIM*jnrA;
 231             j_coord_offsetB  = DIM*jnrB;
 232
 233             /* load j atom coordinates */
 234             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 235                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 236                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 237
 238             /* Calculate displacement vector */
 239             dx00             = _mm_sub_pd(ix0,jx0);
 240             dy00             = _mm_sub_pd(iy0,jy0);
 241             dz00             = _mm_sub_pd(iz0,jz0);
 242             dx11             = _mm_sub_pd(ix1,jx1);
 243             dy11             = _mm_sub_pd(iy1,jy1);
 244             dz11             = _mm_sub_pd(iz1,jz1);
 245             dx12             = _mm_sub_pd(ix1,jx2);
 246             dy12             = _mm_sub_pd(iy1,jy2);
 247             dz12             = _mm_sub_pd(iz1,jz2);
 248             dx13             = _mm_sub_pd(ix1,jx3);
 249             dy13             = _mm_sub_pd(iy1,jy3);
 250             dz13             = _mm_sub_pd(iz1,jz3);
 251             dx21             = _mm_sub_pd(ix2,jx1);
 252             dy21             = _mm_sub_pd(iy2,jy1);
 253             dz21             = _mm_sub_pd(iz2,jz1);
 254             dx22             = _mm_sub_pd(ix2,jx2);
 255             dy22             = _mm_sub_pd(iy2,jy2);
 256             dz22             = _mm_sub_pd(iz2,jz2);
 257             dx23             = _mm_sub_pd(ix2,jx3);
 258             dy23             = _mm_sub_pd(iy2,jy3);
 259             dz23             = _mm_sub_pd(iz2,jz3);
 260             dx31             = _mm_sub_pd(ix3,jx1);
 261             dy31             = _mm_sub_pd(iy3,jy1);
 262             dz31             = _mm_sub_pd(iz3,jz1);
 263             dx32             = _mm_sub_pd(ix3,jx2);
 264             dy32             = _mm_sub_pd(iy3,jy2);
 265             dz32             = _mm_sub_pd(iz3,jz2);
 266             dx33             = _mm_sub_pd(ix3,jx3);
 267             dy33             = _mm_sub_pd(iy3,jy3);
 268             dz33             = _mm_sub_pd(iz3,jz3);
 269
 270             /* Calculate squared distance and things based on it */
 271             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 272             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 273             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 274             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 275             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 276             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 277             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 278             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 279             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 280             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 281
 282             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 283             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 284             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 285             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 286             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 287             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 288             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 289             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 290             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 291             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 292
 293             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 294             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 295             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 296             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 297             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 298             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 299             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 300             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 301             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 302             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 303
 304             fjx0             = _mm_setzero_pd();
 305             fjy0             = _mm_setzero_pd();
 306             fjz0             = _mm_setzero_pd();
 307             fjx1             = _mm_setzero_pd();
 308             fjy1             = _mm_setzero_pd();
 309             fjz1             = _mm_setzero_pd();
 310             fjx2             = _mm_setzero_pd();
 311             fjy2             = _mm_setzero_pd();
 312             fjz2             = _mm_setzero_pd();
 313             fjx3             = _mm_setzero_pd();
 314             fjy3             = _mm_setzero_pd();
 315             fjz3             = _mm_setzero_pd();
 316
 317             /**************************
 318              * CALCULATE INTERACTIONS *
 319              **************************/
 320
 321             if (gmx_mm_any_lt(rsq00,rcutoff2))
 322             {
 323
 324             r00              = _mm_mul_pd(rsq00,rinv00);
 325
 326             /* LENNARD-JONES DISPERSION/REPULSION */
 327
 328             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 329             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 330             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 331             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 332             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 333
 334             d                = _mm_sub_pd(r00,rswitch);
 335             d                = _mm_max_pd(d,_mm_setzero_pd());
 336             d2               = _mm_mul_pd(d,d);
 337             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 338
 339             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 340
 341             /* Evaluate switch function */
 342             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 343             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 344             vvdw             = _mm_mul_pd(vvdw,sw);
 345             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 346
 347             /* Update potential sum for this i atom from the interaction with this j atom. */
 348             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 349             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 350
 351             fscal            = fvdw;
 352
 353             fscal            = _mm_and_pd(fscal,cutoff_mask);
 354
 355             /* Update vectorial force */
 356             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 357             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 358             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 359
 360             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 361             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 362             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 363
 364             }
 365
 366             /**************************
 367              * CALCULATE INTERACTIONS *
 368              **************************/
 369
 370             if (gmx_mm_any_lt(rsq11,rcutoff2))
 371             {
 372
 373             /* REACTION-FIELD ELECTROSTATICS */
 374             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 375             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 376
 377             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 378
 379             /* Update potential sum for this i atom from the interaction with this j atom. */
 380             velec            = _mm_and_pd(velec,cutoff_mask);
 381             velecsum         = _mm_add_pd(velecsum,velec);
 382
 383             fscal            = felec;
 384
 385             fscal            = _mm_and_pd(fscal,cutoff_mask);
 386
 387             /* Update vectorial force */
 388             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 389             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 390             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 391
 392             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 393             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 394             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 395
 396             }
 397
 398             /**************************
 399              * CALCULATE INTERACTIONS *
 400              **************************/
 401
 402             if (gmx_mm_any_lt(rsq12,rcutoff2))
 403             {
 404
 405             /* REACTION-FIELD ELECTROSTATICS */
 406             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 407             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 408
 409             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 410
 411             /* Update potential sum for this i atom from the interaction with this j atom. */
 412             velec            = _mm_and_pd(velec,cutoff_mask);
 413             velecsum         = _mm_add_pd(velecsum,velec);
 414
 415             fscal            = felec;
 416
 417             fscal            = _mm_and_pd(fscal,cutoff_mask);
 418
 419             /* Update vectorial force */
 420             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 421             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 422             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 423
 424             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 425             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 426             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 427
 428             }
 429
 430             /**************************
 431              * CALCULATE INTERACTIONS *
 432              **************************/
 433
 434             if (gmx_mm_any_lt(rsq13,rcutoff2))
 435             {
 436
 437             /* REACTION-FIELD ELECTROSTATICS */
 438             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 439             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 440
 441             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 442
 443             /* Update potential sum for this i atom from the interaction with this j atom. */
 444             velec            = _mm_and_pd(velec,cutoff_mask);
 445             velecsum         = _mm_add_pd(velecsum,velec);
 446
 447             fscal            = felec;
 448
 449             fscal            = _mm_and_pd(fscal,cutoff_mask);
 450
 451             /* Update vectorial force */
 452             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 453             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 454             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 455
 456             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 457             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 458             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 459
 460             }
 461
 462             /**************************
 463              * CALCULATE INTERACTIONS *
 464              **************************/
 465
 466             if (gmx_mm_any_lt(rsq21,rcutoff2))
 467             {
 468
 469             /* REACTION-FIELD ELECTROSTATICS */
 470             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 471             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 472
 473             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 474
 475             /* Update potential sum for this i atom from the interaction with this j atom. */
 476             velec            = _mm_and_pd(velec,cutoff_mask);
 477             velecsum         = _mm_add_pd(velecsum,velec);
 478
 479             fscal            = felec;
 480
 481             fscal            = _mm_and_pd(fscal,cutoff_mask);
 482
 483             /* Update vectorial force */
 484             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 485             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 486             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 487
 488             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 489             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 490             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 491
 492             }
 493
 494             /**************************
 495              * CALCULATE INTERACTIONS *
 496              **************************/
 497
 498             if (gmx_mm_any_lt(rsq22,rcutoff2))
 499             {
 500
 501             /* REACTION-FIELD ELECTROSTATICS */
 502             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 503             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 504
 505             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 506
 507             /* Update potential sum for this i atom from the interaction with this j atom. */
 508             velec            = _mm_and_pd(velec,cutoff_mask);
 509             velecsum         = _mm_add_pd(velecsum,velec);
 510
 511             fscal            = felec;
 512
 513             fscal            = _mm_and_pd(fscal,cutoff_mask);
 514
 515             /* Update vectorial force */
 516             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 517             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 518             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 519
 520             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 521             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 522             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 523
 524             }
 525
 526             /**************************
 527              * CALCULATE INTERACTIONS *
 528              **************************/
 529
 530             if (gmx_mm_any_lt(rsq23,rcutoff2))
 531             {
 532
 533             /* REACTION-FIELD ELECTROSTATICS */
 534             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 535             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 536
 537             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 538
 539             /* Update potential sum for this i atom from the interaction with this j atom. */
 540             velec            = _mm_and_pd(velec,cutoff_mask);
 541             velecsum         = _mm_add_pd(velecsum,velec);
 542
 543             fscal            = felec;
 544
 545             fscal            = _mm_and_pd(fscal,cutoff_mask);
 546
 547             /* Update vectorial force */
 548             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 549             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 550             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 551
 552             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 553             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 554             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 555
 556             }
 557
 558             /**************************
 559              * CALCULATE INTERACTIONS *
 560              **************************/
 561
 562             if (gmx_mm_any_lt(rsq31,rcutoff2))
 563             {
 564
 565             /* REACTION-FIELD ELECTROSTATICS */
 566             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
 567             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
 568
 569             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
 570
 571             /* Update potential sum for this i atom from the interaction with this j atom. */
 572             velec            = _mm_and_pd(velec,cutoff_mask);
 573             velecsum         = _mm_add_pd(velecsum,velec);
 574
 575             fscal            = felec;
 576
 577             fscal            = _mm_and_pd(fscal,cutoff_mask);
 578
 579             /* Update vectorial force */
 580             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 581             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 582             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 583
 584             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 585             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 586             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 587
 588             }
 589
 590             /**************************
 591              * CALCULATE INTERACTIONS *
 592              **************************/
 593
 594             if (gmx_mm_any_lt(rsq32,rcutoff2))
 595             {
 596
 597             /* REACTION-FIELD ELECTROSTATICS */
 598             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
 599             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
 600
 601             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
 602
 603             /* Update potential sum for this i atom from the interaction with this j atom. */
 604             velec            = _mm_and_pd(velec,cutoff_mask);
 605             velecsum         = _mm_add_pd(velecsum,velec);
 606
 607             fscal            = felec;
 608
 609             fscal            = _mm_and_pd(fscal,cutoff_mask);
 610
 611             /* Update vectorial force */
 612             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 613             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 614             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 615
 616             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 617             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 618             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 619
 620             }
 621
 622             /**************************
 623              * CALCULATE INTERACTIONS *
 624              **************************/
 625
 626             if (gmx_mm_any_lt(rsq33,rcutoff2))
 627             {
 628
 629             /* REACTION-FIELD ELECTROSTATICS */
 630             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
 631             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
 632
 633             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
 634
 635             /* Update potential sum for this i atom from the interaction with this j atom. */
 636             velec            = _mm_and_pd(velec,cutoff_mask);
 637             velecsum         = _mm_add_pd(velecsum,velec);
 638
 639             fscal            = felec;
 640
 641             fscal            = _mm_and_pd(fscal,cutoff_mask);
 642
 643             /* Update vectorial force */
 644             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 645             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 646             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 647
 648             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 649             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 650             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 651
 652             }
 653
 654             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);
 655
 656             /* Inner loop uses 416 flops */
 657         }
 658
 659         if(jidx<j_index_end)
 660         {
 661
 662             jnrA             = jjnr[jidx];
 663             j_coord_offsetA  = DIM*jnrA;
 664
 665             /* load j atom coordinates */
 666             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 667                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 668                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 669
 670             /* Calculate displacement vector */
 671             dx00             = _mm_sub_pd(ix0,jx0);
 672             dy00             = _mm_sub_pd(iy0,jy0);
 673             dz00             = _mm_sub_pd(iz0,jz0);
 674             dx11             = _mm_sub_pd(ix1,jx1);
 675             dy11             = _mm_sub_pd(iy1,jy1);
 676             dz11             = _mm_sub_pd(iz1,jz1);
 677             dx12             = _mm_sub_pd(ix1,jx2);
 678             dy12             = _mm_sub_pd(iy1,jy2);
 679             dz12             = _mm_sub_pd(iz1,jz2);
 680             dx13             = _mm_sub_pd(ix1,jx3);
 681             dy13             = _mm_sub_pd(iy1,jy3);
 682             dz13             = _mm_sub_pd(iz1,jz3);
 683             dx21             = _mm_sub_pd(ix2,jx1);
 684             dy21             = _mm_sub_pd(iy2,jy1);
 685             dz21             = _mm_sub_pd(iz2,jz1);
 686             dx22             = _mm_sub_pd(ix2,jx2);
 687             dy22             = _mm_sub_pd(iy2,jy2);
 688             dz22             = _mm_sub_pd(iz2,jz2);
 689             dx23             = _mm_sub_pd(ix2,jx3);
 690             dy23             = _mm_sub_pd(iy2,jy3);
 691             dz23             = _mm_sub_pd(iz2,jz3);
 692             dx31             = _mm_sub_pd(ix3,jx1);
 693             dy31             = _mm_sub_pd(iy3,jy1);
 694             dz31             = _mm_sub_pd(iz3,jz1);
 695             dx32             = _mm_sub_pd(ix3,jx2);
 696             dy32             = _mm_sub_pd(iy3,jy2);
 697             dz32             = _mm_sub_pd(iz3,jz2);
 698             dx33             = _mm_sub_pd(ix3,jx3);
 699             dy33             = _mm_sub_pd(iy3,jy3);
 700             dz33             = _mm_sub_pd(iz3,jz3);
 701
 702             /* Calculate squared distance and things based on it */
 703             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 704             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 705             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 706             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 707             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 708             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 709             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 710             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 711             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 712             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 713
 714             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 715             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 716             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 717             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 718             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 719             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 720             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 721             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 722             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 723             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 724
 725             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 726             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 727             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 728             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 729             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 730             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 731             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 732             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 733             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 734             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 735
 736             fjx0             = _mm_setzero_pd();
 737             fjy0             = _mm_setzero_pd();
 738             fjz0             = _mm_setzero_pd();
 739             fjx1             = _mm_setzero_pd();
 740             fjy1             = _mm_setzero_pd();
 741             fjz1             = _mm_setzero_pd();
 742             fjx2             = _mm_setzero_pd();
 743             fjy2             = _mm_setzero_pd();
 744             fjz2             = _mm_setzero_pd();
 745             fjx3             = _mm_setzero_pd();
 746             fjy3             = _mm_setzero_pd();
 747             fjz3             = _mm_setzero_pd();
 748
 749             /**************************
 750              * CALCULATE INTERACTIONS *
 751              **************************/
 752
 753             if (gmx_mm_any_lt(rsq00,rcutoff2))
 754             {
 755
 756             r00              = _mm_mul_pd(rsq00,rinv00);
 757
 758             /* LENNARD-JONES DISPERSION/REPULSION */
 759
 760             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 761             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 762             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 763             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 764             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 765
 766             d                = _mm_sub_pd(r00,rswitch);
 767             d                = _mm_max_pd(d,_mm_setzero_pd());
 768             d2               = _mm_mul_pd(d,d);
 769             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 770
 771             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 772
 773             /* Evaluate switch function */
 774             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 775             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 776             vvdw             = _mm_mul_pd(vvdw,sw);
 777             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 778
 779             /* Update potential sum for this i atom from the interaction with this j atom. */
 780             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 781             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 782             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 783
 784             fscal            = fvdw;
 785
 786             fscal            = _mm_and_pd(fscal,cutoff_mask);
 787
 788             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 789
 790             /* Update vectorial force */
 791             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 792             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 793             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 794
 795             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 796             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 797             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 798
 799             }
 800
 801             /**************************
 802              * CALCULATE INTERACTIONS *
 803              **************************/
 804
 805             if (gmx_mm_any_lt(rsq11,rcutoff2))
 806             {
 807
 808             /* REACTION-FIELD ELECTROSTATICS */
 809             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 810             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 811
 812             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 813
 814             /* Update potential sum for this i atom from the interaction with this j atom. */
 815             velec            = _mm_and_pd(velec,cutoff_mask);
 816             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 817             velecsum         = _mm_add_pd(velecsum,velec);
 818
 819             fscal            = felec;
 820
 821             fscal            = _mm_and_pd(fscal,cutoff_mask);
 822
 823             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 824
 825             /* Update vectorial force */
 826             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 827             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 828             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 829
 830             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 831             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 832             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 833
 834             }
 835
 836             /**************************
 837              * CALCULATE INTERACTIONS *
 838              **************************/
 839
 840             if (gmx_mm_any_lt(rsq12,rcutoff2))
 841             {
 842
 843             /* REACTION-FIELD ELECTROSTATICS */
 844             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 845             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 846
 847             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 848
 849             /* Update potential sum for this i atom from the interaction with this j atom. */
 850             velec            = _mm_and_pd(velec,cutoff_mask);
 851             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 852             velecsum         = _mm_add_pd(velecsum,velec);
 853
 854             fscal            = felec;
 855
 856             fscal            = _mm_and_pd(fscal,cutoff_mask);
 857
 858             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 859
 860             /* Update vectorial force */
 861             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 862             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 863             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 864
 865             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 866             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 867             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 868
 869             }
 870
 871             /**************************
 872              * CALCULATE INTERACTIONS *
 873              **************************/
 874
 875             if (gmx_mm_any_lt(rsq13,rcutoff2))
 876             {
 877
 878             /* REACTION-FIELD ELECTROSTATICS */
 879             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 880             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 881
 882             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 883
 884             /* Update potential sum for this i atom from the interaction with this j atom. */
 885             velec            = _mm_and_pd(velec,cutoff_mask);
 886             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 887             velecsum         = _mm_add_pd(velecsum,velec);
 888
 889             fscal            = felec;
 890
 891             fscal            = _mm_and_pd(fscal,cutoff_mask);
 892
 893             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 894
 895             /* Update vectorial force */
 896             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 897             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 898             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 899
 900             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 901             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 902             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 903
 904             }
 905
 906             /**************************
 907              * CALCULATE INTERACTIONS *
 908              **************************/
 909
 910             if (gmx_mm_any_lt(rsq21,rcutoff2))
 911             {
 912
 913             /* REACTION-FIELD ELECTROSTATICS */
 914             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 915             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 916
 917             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 918
 919             /* Update potential sum for this i atom from the interaction with this j atom. */
 920             velec            = _mm_and_pd(velec,cutoff_mask);
 921             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 922             velecsum         = _mm_add_pd(velecsum,velec);
 923
 924             fscal            = felec;
 925
 926             fscal            = _mm_and_pd(fscal,cutoff_mask);
 927
 928             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 929
 930             /* Update vectorial force */
 931             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 932             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 933             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 934
 935             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 936             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 937             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 938
 939             }
 940
 941             /**************************
 942              * CALCULATE INTERACTIONS *
 943              **************************/
 944
 945             if (gmx_mm_any_lt(rsq22,rcutoff2))
 946             {
 947
 948             /* REACTION-FIELD ELECTROSTATICS */
 949             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 950             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 951
 952             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 953
 954             /* Update potential sum for this i atom from the interaction with this j atom. */
 955             velec            = _mm_and_pd(velec,cutoff_mask);
 956             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 957             velecsum         = _mm_add_pd(velecsum,velec);
 958
 959             fscal            = felec;
 960
 961             fscal            = _mm_and_pd(fscal,cutoff_mask);
 962
 963             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 964
 965             /* Update vectorial force */
 966             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 967             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 968             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 969
 970             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 971             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 972             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 973
 974             }
 975
 976             /**************************
 977              * CALCULATE INTERACTIONS *
 978              **************************/
 979
 980             if (gmx_mm_any_lt(rsq23,rcutoff2))
 981             {
 982
 983             /* REACTION-FIELD ELECTROSTATICS */
 984             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 985             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 986
 987             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 988
 989             /* Update potential sum for this i atom from the interaction with this j atom. */
 990             velec            = _mm_and_pd(velec,cutoff_mask);
 991             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 992             velecsum         = _mm_add_pd(velecsum,velec);
 993
 994             fscal            = felec;
 995
 996             fscal            = _mm_and_pd(fscal,cutoff_mask);
 997
 998             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 999
1000             /* Update vectorial force */
1001             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1002             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1003             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1004
1005             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1006             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1007             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1008
1009             }
1010
1011             /**************************
1012              * CALCULATE INTERACTIONS *
1013              **************************/
1014
1015             if (gmx_mm_any_lt(rsq31,rcutoff2))
1016             {
1017
1018             /* REACTION-FIELD ELECTROSTATICS */
1019             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
1020             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1021
1022             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1023
1024             /* Update potential sum for this i atom from the interaction with this j atom. */
1025             velec            = _mm_and_pd(velec,cutoff_mask);
1026             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1027             velecsum         = _mm_add_pd(velecsum,velec);
1028
1029             fscal            = felec;
1030
1031             fscal            = _mm_and_pd(fscal,cutoff_mask);
1032
1033             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1034
1035             /* Update vectorial force */
1036             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1037             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1038             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1039
1040             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1041             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1042             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1043
1044             }
1045
1046             /**************************
1047              * CALCULATE INTERACTIONS *
1048              **************************/
1049
1050             if (gmx_mm_any_lt(rsq32,rcutoff2))
1051             {
1052
1053             /* REACTION-FIELD ELECTROSTATICS */
1054             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
1055             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1056
1057             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1058
1059             /* Update potential sum for this i atom from the interaction with this j atom. */
1060             velec            = _mm_and_pd(velec,cutoff_mask);
1061             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1062             velecsum         = _mm_add_pd(velecsum,velec);
1063
1064             fscal            = felec;
1065
1066             fscal            = _mm_and_pd(fscal,cutoff_mask);
1067
1068             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1069
1070             /* Update vectorial force */
1071             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1072             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1073             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1074
1075             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1076             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1077             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1078
1079             }
1080
1081             /**************************
1082              * CALCULATE INTERACTIONS *
1083              **************************/
1084
1085             if (gmx_mm_any_lt(rsq33,rcutoff2))
1086             {
1087
1088             /* REACTION-FIELD ELECTROSTATICS */
1089             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
1090             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1091
1092             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1093
1094             /* Update potential sum for this i atom from the interaction with this j atom. */
1095             velec            = _mm_and_pd(velec,cutoff_mask);
1096             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1097             velecsum         = _mm_add_pd(velecsum,velec);
1098
1099             fscal            = felec;
1100
1101             fscal            = _mm_and_pd(fscal,cutoff_mask);
1102
1103             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1104
1105             /* Update vectorial force */
1106             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1107             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1108             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1109
1110             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1111             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1112             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1113
1114             }
1115
1116             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1117
1118             /* Inner loop uses 416 flops */
1119         }
1120
1121         /* End of innermost loop */
1122
1123         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1124                                               f+i_coord_offset,fshift+i_shift_offset);
1125
1126         ggid                        = gid[iidx];
1127         /* Update potential energies */
1128         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1129         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
1130
1131         /* Increment number of inner iterations */
1132         inneriter                  += j_index_end - j_index_start;
1133
1134         /* Outer loop uses 26 flops */
1135     }
1136
1137     /* Increment number of outer iterations */
1138     outeriter        += nri;
1139
1140     /* Update outer/inner flops */
1141
1142     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*416);
1143 }
1144 /*
1145  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_F_avx_128_fma_double
1146  * Electrostatics interaction: ReactionField
1147  * VdW interaction:            LennardJones
1148  * Geometry:                   Water4-Water4
1149  * Calculate force/pot:        Force
1150  */
1151 void
1152 nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_F_avx_128_fma_double
1153                     (t_nblist                    * gmx_restrict       nlist,
1154                      rvec                        * gmx_restrict          xx,
1155                      rvec                        * gmx_restrict          ff,
1156                      t_forcerec                  * gmx_restrict          fr,
1157                      t_mdatoms                   * gmx_restrict     mdatoms,
1158                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1159                      t_nrnb                      * gmx_restrict        nrnb)
1160 {
1161     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1162      * just 0 for non-waters.
1163      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1164      * jnr indices corresponding to data put in the four positions in the SIMD register.
1165      */
1166     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1167     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1168     int              jnrA,jnrB;
1169     int              j_coord_offsetA,j_coord_offsetB;
1170     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1171     real             rcutoff_scalar;
1172     real             *shiftvec,*fshift,*x,*f;
1173     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1174     int              vdwioffset0;
1175     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1176     int              vdwioffset1;
1177     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1178     int              vdwioffset2;
1179     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1180     int              vdwioffset3;
1181     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1182     int              vdwjidx0A,vdwjidx0B;
1183     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1184     int              vdwjidx1A,vdwjidx1B;
1185     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1186     int              vdwjidx2A,vdwjidx2B;
1187     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1188     int              vdwjidx3A,vdwjidx3B;
1189     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1190     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1191     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1192     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1193     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1194     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1195     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1196     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1197     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1198     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1199     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1200     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1201     real             *charge;
1202     int              nvdwtype;
1203     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1204     int              *vdwtype;
1205     real             *vdwparam;
1206     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1207     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1208     __m128d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
1209     real             rswitch_scalar,d_scalar;
1210     __m128d          dummy_mask,cutoff_mask;
1211     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1212     __m128d          one     = _mm_set1_pd(1.0);
1213     __m128d          two     = _mm_set1_pd(2.0);
1214     x                = xx[0];
1215     f                = ff[0];
1216
1217     nri              = nlist->nri;
1218     iinr             = nlist->iinr;
1219     jindex           = nlist->jindex;
1220     jjnr             = nlist->jjnr;
1221     shiftidx         = nlist->shift;
1222     gid              = nlist->gid;
1223     shiftvec         = fr->shift_vec[0];
1224     fshift           = fr->fshift[0];
1225     facel            = _mm_set1_pd(fr->epsfac);
1226     charge           = mdatoms->chargeA;
1227     krf              = _mm_set1_pd(fr->ic->k_rf);
1228     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1229     crf              = _mm_set1_pd(fr->ic->c_rf);
1230     nvdwtype         = fr->ntype;
1231     vdwparam         = fr->nbfp;
1232     vdwtype          = mdatoms->typeA;
1233
1234     /* Setup water-specific parameters */
1235     inr              = nlist->iinr[0];
1236     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1237     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1238     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1239     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1240
1241     jq1              = _mm_set1_pd(charge[inr+1]);
1242     jq2              = _mm_set1_pd(charge[inr+2]);
1243     jq3              = _mm_set1_pd(charge[inr+3]);
1244     vdwjidx0A        = 2*vdwtype[inr+0];
1245     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1246     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1247     qq11             = _mm_mul_pd(iq1,jq1);
1248     qq12             = _mm_mul_pd(iq1,jq2);
1249     qq13             = _mm_mul_pd(iq1,jq3);
1250     qq21             = _mm_mul_pd(iq2,jq1);
1251     qq22             = _mm_mul_pd(iq2,jq2);
1252     qq23             = _mm_mul_pd(iq2,jq3);
1253     qq31             = _mm_mul_pd(iq3,jq1);
1254     qq32             = _mm_mul_pd(iq3,jq2);
1255     qq33             = _mm_mul_pd(iq3,jq3);
1256
1257     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1258     rcutoff_scalar   = fr->rcoulomb;
1259     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1260     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1261
1262     rswitch_scalar   = fr->rvdw_switch;
1263     rswitch          = _mm_set1_pd(rswitch_scalar);
1264     /* Setup switch parameters */
1265     d_scalar         = rcutoff_scalar-rswitch_scalar;
1266     d                = _mm_set1_pd(d_scalar);
1267     swV3             = _mm_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
1268     swV4             = _mm_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1269     swV5             = _mm_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1270     swF2             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
1271     swF3             = _mm_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1272     swF4             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1273
1274     /* Avoid stupid compiler warnings */
1275     jnrA = jnrB = 0;
1276     j_coord_offsetA = 0;
1277     j_coord_offsetB = 0;
1278
1279     outeriter        = 0;
1280     inneriter        = 0;
1281
1282     /* Start outer loop over neighborlists */
1283     for(iidx=0; iidx<nri; iidx++)
1284     {
1285         /* Load shift vector for this list */
1286         i_shift_offset   = DIM*shiftidx[iidx];
1287
1288         /* Load limits for loop over neighbors */
1289         j_index_start    = jindex[iidx];
1290         j_index_end      = jindex[iidx+1];
1291
1292         /* Get outer coordinate index */
1293         inr              = iinr[iidx];
1294         i_coord_offset   = DIM*inr;
1295
1296         /* Load i particle coords and add shift vector */
1297         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1298                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1299
1300         fix0             = _mm_setzero_pd();
1301         fiy0             = _mm_setzero_pd();
1302         fiz0             = _mm_setzero_pd();
1303         fix1             = _mm_setzero_pd();
1304         fiy1             = _mm_setzero_pd();
1305         fiz1             = _mm_setzero_pd();
1306         fix2             = _mm_setzero_pd();
1307         fiy2             = _mm_setzero_pd();
1308         fiz2             = _mm_setzero_pd();
1309         fix3             = _mm_setzero_pd();
1310         fiy3             = _mm_setzero_pd();
1311         fiz3             = _mm_setzero_pd();
1312
1313         /* Start inner kernel loop */
1314         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1315         {
1316
1317             /* Get j neighbor index, and coordinate index */
1318             jnrA             = jjnr[jidx];
1319             jnrB             = jjnr[jidx+1];
1320             j_coord_offsetA  = DIM*jnrA;
1321             j_coord_offsetB  = DIM*jnrB;
1322
1323             /* load j atom coordinates */
1324             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1325                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1326                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1327
1328             /* Calculate displacement vector */
1329             dx00             = _mm_sub_pd(ix0,jx0);
1330             dy00             = _mm_sub_pd(iy0,jy0);
1331             dz00             = _mm_sub_pd(iz0,jz0);
1332             dx11             = _mm_sub_pd(ix1,jx1);
1333             dy11             = _mm_sub_pd(iy1,jy1);
1334             dz11             = _mm_sub_pd(iz1,jz1);
1335             dx12             = _mm_sub_pd(ix1,jx2);
1336             dy12             = _mm_sub_pd(iy1,jy2);
1337             dz12             = _mm_sub_pd(iz1,jz2);
1338             dx13             = _mm_sub_pd(ix1,jx3);
1339             dy13             = _mm_sub_pd(iy1,jy3);
1340             dz13             = _mm_sub_pd(iz1,jz3);
1341             dx21             = _mm_sub_pd(ix2,jx1);
1342             dy21             = _mm_sub_pd(iy2,jy1);
1343             dz21             = _mm_sub_pd(iz2,jz1);
1344             dx22             = _mm_sub_pd(ix2,jx2);
1345             dy22             = _mm_sub_pd(iy2,jy2);
1346             dz22             = _mm_sub_pd(iz2,jz2);
1347             dx23             = _mm_sub_pd(ix2,jx3);
1348             dy23             = _mm_sub_pd(iy2,jy3);
1349             dz23             = _mm_sub_pd(iz2,jz3);
1350             dx31             = _mm_sub_pd(ix3,jx1);
1351             dy31             = _mm_sub_pd(iy3,jy1);
1352             dz31             = _mm_sub_pd(iz3,jz1);
1353             dx32             = _mm_sub_pd(ix3,jx2);
1354             dy32             = _mm_sub_pd(iy3,jy2);
1355             dz32             = _mm_sub_pd(iz3,jz2);
1356             dx33             = _mm_sub_pd(ix3,jx3);
1357             dy33             = _mm_sub_pd(iy3,jy3);
1358             dz33             = _mm_sub_pd(iz3,jz3);
1359
1360             /* Calculate squared distance and things based on it */
1361             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1362             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1363             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1364             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1365             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1366             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1367             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1368             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1369             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1370             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1371
1372             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1373             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1374             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1375             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1376             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1377             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1378             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1379             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1380             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1381             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1382
1383             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1384             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1385             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1386             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1387             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1388             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1389             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1390             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1391             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1392             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1393
1394             fjx0             = _mm_setzero_pd();
1395             fjy0             = _mm_setzero_pd();
1396             fjz0             = _mm_setzero_pd();
1397             fjx1             = _mm_setzero_pd();
1398             fjy1             = _mm_setzero_pd();
1399             fjz1             = _mm_setzero_pd();
1400             fjx2             = _mm_setzero_pd();
1401             fjy2             = _mm_setzero_pd();
1402             fjz2             = _mm_setzero_pd();
1403             fjx3             = _mm_setzero_pd();
1404             fjy3             = _mm_setzero_pd();
1405             fjz3             = _mm_setzero_pd();
1406
1407             /**************************
1408              * CALCULATE INTERACTIONS *
1409              **************************/
1410
1411             if (gmx_mm_any_lt(rsq00,rcutoff2))
1412             {
1413
1414             r00              = _mm_mul_pd(rsq00,rinv00);
1415
1416             /* LENNARD-JONES DISPERSION/REPULSION */
1417
1418             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1419             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
1420             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
1421             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
1422             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
1423
1424             d                = _mm_sub_pd(r00,rswitch);
1425             d                = _mm_max_pd(d,_mm_setzero_pd());
1426             d2               = _mm_mul_pd(d,d);
1427             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
1428
1429             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
1430
1431             /* Evaluate switch function */
1432             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1433             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
1434             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1435
1436             fscal            = fvdw;
1437
1438             fscal            = _mm_and_pd(fscal,cutoff_mask);
1439
1440             /* Update vectorial force */
1441             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1442             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1443             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1444
1445             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1446             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1447             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1448
1449             }
1450
1451             /**************************
1452              * CALCULATE INTERACTIONS *
1453              **************************/
1454
1455             if (gmx_mm_any_lt(rsq11,rcutoff2))
1456             {
1457
1458             /* REACTION-FIELD ELECTROSTATICS */
1459             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1460
1461             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1462
1463             fscal            = felec;
1464
1465             fscal            = _mm_and_pd(fscal,cutoff_mask);
1466
1467             /* Update vectorial force */
1468             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1469             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1470             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1471
1472             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1473             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1474             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1475
1476             }
1477
1478             /**************************
1479              * CALCULATE INTERACTIONS *
1480              **************************/
1481
1482             if (gmx_mm_any_lt(rsq12,rcutoff2))
1483             {
1484
1485             /* REACTION-FIELD ELECTROSTATICS */
1486             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1487
1488             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1489
1490             fscal            = felec;
1491
1492             fscal            = _mm_and_pd(fscal,cutoff_mask);
1493
1494             /* Update vectorial force */
1495             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1496             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1497             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1498
1499             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1500             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1501             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1502
1503             }
1504
1505             /**************************
1506              * CALCULATE INTERACTIONS *
1507              **************************/
1508
1509             if (gmx_mm_any_lt(rsq13,rcutoff2))
1510             {
1511
1512             /* REACTION-FIELD ELECTROSTATICS */
1513             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1514
1515             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1516
1517             fscal            = felec;
1518
1519             fscal            = _mm_and_pd(fscal,cutoff_mask);
1520
1521             /* Update vectorial force */
1522             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1523             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1524             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1525
1526             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1527             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1528             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1529
1530             }
1531
1532             /**************************
1533              * CALCULATE INTERACTIONS *
1534              **************************/
1535
1536             if (gmx_mm_any_lt(rsq21,rcutoff2))
1537             {
1538
1539             /* REACTION-FIELD ELECTROSTATICS */
1540             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1541
1542             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1543
1544             fscal            = felec;
1545
1546             fscal            = _mm_and_pd(fscal,cutoff_mask);
1547
1548             /* Update vectorial force */
1549             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1550             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1551             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1552
1553             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1554             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1555             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1556
1557             }
1558
1559             /**************************
1560              * CALCULATE INTERACTIONS *
1561              **************************/
1562
1563             if (gmx_mm_any_lt(rsq22,rcutoff2))
1564             {
1565
1566             /* REACTION-FIELD ELECTROSTATICS */
1567             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1568
1569             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1570
1571             fscal            = felec;
1572
1573             fscal            = _mm_and_pd(fscal,cutoff_mask);
1574
1575             /* Update vectorial force */
1576             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1577             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1578             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1579
1580             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1581             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1582             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1583
1584             }
1585
1586             /**************************
1587              * CALCULATE INTERACTIONS *
1588              **************************/
1589
1590             if (gmx_mm_any_lt(rsq23,rcutoff2))
1591             {
1592
1593             /* REACTION-FIELD ELECTROSTATICS */
1594             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1595
1596             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1597
1598             fscal            = felec;
1599
1600             fscal            = _mm_and_pd(fscal,cutoff_mask);
1601
1602             /* Update vectorial force */
1603             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1604             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1605             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1606
1607             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1608             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1609             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1610
1611             }
1612
1613             /**************************
1614              * CALCULATE INTERACTIONS *
1615              **************************/
1616
1617             if (gmx_mm_any_lt(rsq31,rcutoff2))
1618             {
1619
1620             /* REACTION-FIELD ELECTROSTATICS */
1621             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1622
1623             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1624
1625             fscal            = felec;
1626
1627             fscal            = _mm_and_pd(fscal,cutoff_mask);
1628
1629             /* Update vectorial force */
1630             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1631             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1632             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1633
1634             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1635             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1636             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1637
1638             }
1639
1640             /**************************
1641              * CALCULATE INTERACTIONS *
1642              **************************/
1643
1644             if (gmx_mm_any_lt(rsq32,rcutoff2))
1645             {
1646
1647             /* REACTION-FIELD ELECTROSTATICS */
1648             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1649
1650             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1651
1652             fscal            = felec;
1653
1654             fscal            = _mm_and_pd(fscal,cutoff_mask);
1655
1656             /* Update vectorial force */
1657             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1658             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1659             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1660
1661             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1662             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1663             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1664
1665             }
1666
1667             /**************************
1668              * CALCULATE INTERACTIONS *
1669              **************************/
1670
1671             if (gmx_mm_any_lt(rsq33,rcutoff2))
1672             {
1673
1674             /* REACTION-FIELD ELECTROSTATICS */
1675             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1676
1677             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1678
1679             fscal            = felec;
1680
1681             fscal            = _mm_and_pd(fscal,cutoff_mask);
1682
1683             /* Update vectorial force */
1684             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1685             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1686             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1687
1688             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1689             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1690             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1691
1692             }
1693
1694             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);
1695
1696             /* Inner loop uses 359 flops */
1697         }
1698
1699         if(jidx<j_index_end)
1700         {
1701
1702             jnrA             = jjnr[jidx];
1703             j_coord_offsetA  = DIM*jnrA;
1704
1705             /* load j atom coordinates */
1706             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1707                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1708                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1709
1710             /* Calculate displacement vector */
1711             dx00             = _mm_sub_pd(ix0,jx0);
1712             dy00             = _mm_sub_pd(iy0,jy0);
1713             dz00             = _mm_sub_pd(iz0,jz0);
1714             dx11             = _mm_sub_pd(ix1,jx1);
1715             dy11             = _mm_sub_pd(iy1,jy1);
1716             dz11             = _mm_sub_pd(iz1,jz1);
1717             dx12             = _mm_sub_pd(ix1,jx2);
1718             dy12             = _mm_sub_pd(iy1,jy2);
1719             dz12             = _mm_sub_pd(iz1,jz2);
1720             dx13             = _mm_sub_pd(ix1,jx3);
1721             dy13             = _mm_sub_pd(iy1,jy3);
1722             dz13             = _mm_sub_pd(iz1,jz3);
1723             dx21             = _mm_sub_pd(ix2,jx1);
1724             dy21             = _mm_sub_pd(iy2,jy1);
1725             dz21             = _mm_sub_pd(iz2,jz1);
1726             dx22             = _mm_sub_pd(ix2,jx2);
1727             dy22             = _mm_sub_pd(iy2,jy2);
1728             dz22             = _mm_sub_pd(iz2,jz2);
1729             dx23             = _mm_sub_pd(ix2,jx3);
1730             dy23             = _mm_sub_pd(iy2,jy3);
1731             dz23             = _mm_sub_pd(iz2,jz3);
1732             dx31             = _mm_sub_pd(ix3,jx1);
1733             dy31             = _mm_sub_pd(iy3,jy1);
1734             dz31             = _mm_sub_pd(iz3,jz1);
1735             dx32             = _mm_sub_pd(ix3,jx2);
1736             dy32             = _mm_sub_pd(iy3,jy2);
1737             dz32             = _mm_sub_pd(iz3,jz2);
1738             dx33             = _mm_sub_pd(ix3,jx3);
1739             dy33             = _mm_sub_pd(iy3,jy3);
1740             dz33             = _mm_sub_pd(iz3,jz3);
1741
1742             /* Calculate squared distance and things based on it */
1743             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1744             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1745             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1746             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1747             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1748             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1749             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1750             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1751             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1752             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1753
1754             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1755             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1756             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1757             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1758             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1759             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1760             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1761             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1762             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1763             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1764
1765             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1766             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1767             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1768             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1769             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1770             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1771             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1772             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1773             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1774             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1775
1776             fjx0             = _mm_setzero_pd();
1777             fjy0             = _mm_setzero_pd();
1778             fjz0             = _mm_setzero_pd();
1779             fjx1             = _mm_setzero_pd();
1780             fjy1             = _mm_setzero_pd();
1781             fjz1             = _mm_setzero_pd();
1782             fjx2             = _mm_setzero_pd();
1783             fjy2             = _mm_setzero_pd();
1784             fjz2             = _mm_setzero_pd();
1785             fjx3             = _mm_setzero_pd();
1786             fjy3             = _mm_setzero_pd();
1787             fjz3             = _mm_setzero_pd();
1788
1789             /**************************
1790              * CALCULATE INTERACTIONS *
1791              **************************/
1792
1793             if (gmx_mm_any_lt(rsq00,rcutoff2))
1794             {
1795
1796             r00              = _mm_mul_pd(rsq00,rinv00);
1797
1798             /* LENNARD-JONES DISPERSION/REPULSION */
1799
1800             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1801             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
1802             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
1803             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
1804             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
1805
1806             d                = _mm_sub_pd(r00,rswitch);
1807             d                = _mm_max_pd(d,_mm_setzero_pd());
1808             d2               = _mm_mul_pd(d,d);
1809             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
1810
1811             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
1812
1813             /* Evaluate switch function */
1814             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1815             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
1816             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1817
1818             fscal            = fvdw;
1819
1820             fscal            = _mm_and_pd(fscal,cutoff_mask);
1821
1822             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1823
1824             /* Update vectorial force */
1825             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1826             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1827             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1828
1829             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1830             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1831             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1832
1833             }
1834
1835             /**************************
1836              * CALCULATE INTERACTIONS *
1837              **************************/
1838
1839             if (gmx_mm_any_lt(rsq11,rcutoff2))
1840             {
1841
1842             /* REACTION-FIELD ELECTROSTATICS */
1843             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1844
1845             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1846
1847             fscal            = felec;
1848
1849             fscal            = _mm_and_pd(fscal,cutoff_mask);
1850
1851             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1852
1853             /* Update vectorial force */
1854             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1855             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1856             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1857
1858             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1859             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1860             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1861
1862             }
1863
1864             /**************************
1865              * CALCULATE INTERACTIONS *
1866              **************************/
1867
1868             if (gmx_mm_any_lt(rsq12,rcutoff2))
1869             {
1870
1871             /* REACTION-FIELD ELECTROSTATICS */
1872             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1873
1874             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1875
1876             fscal            = felec;
1877
1878             fscal            = _mm_and_pd(fscal,cutoff_mask);
1879
1880             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1881
1882             /* Update vectorial force */
1883             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1884             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1885             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1886
1887             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1888             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1889             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1890
1891             }
1892
1893             /**************************
1894              * CALCULATE INTERACTIONS *
1895              **************************/
1896
1897             if (gmx_mm_any_lt(rsq13,rcutoff2))
1898             {
1899
1900             /* REACTION-FIELD ELECTROSTATICS */
1901             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1902
1903             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1904
1905             fscal            = felec;
1906
1907             fscal            = _mm_and_pd(fscal,cutoff_mask);
1908
1909             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1910
1911             /* Update vectorial force */
1912             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1913             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1914             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1915
1916             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1917             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1918             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1919
1920             }
1921
1922             /**************************
1923              * CALCULATE INTERACTIONS *
1924              **************************/
1925
1926             if (gmx_mm_any_lt(rsq21,rcutoff2))
1927             {
1928
1929             /* REACTION-FIELD ELECTROSTATICS */
1930             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1931
1932             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1933
1934             fscal            = felec;
1935
1936             fscal            = _mm_and_pd(fscal,cutoff_mask);
1937
1938             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1939
1940             /* Update vectorial force */
1941             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1942             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1943             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1944
1945             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1946             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1947             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1948
1949             }
1950
1951             /**************************
1952              * CALCULATE INTERACTIONS *
1953              **************************/
1954
1955             if (gmx_mm_any_lt(rsq22,rcutoff2))
1956             {
1957
1958             /* REACTION-FIELD ELECTROSTATICS */
1959             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1960
1961             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1962
1963             fscal            = felec;
1964
1965             fscal            = _mm_and_pd(fscal,cutoff_mask);
1966
1967             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1968
1969             /* Update vectorial force */
1970             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1971             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1972             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1973
1974             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1975             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1976             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1977
1978             }
1979
1980             /**************************
1981              * CALCULATE INTERACTIONS *
1982              **************************/
1983
1984             if (gmx_mm_any_lt(rsq23,rcutoff2))
1985             {
1986
1987             /* REACTION-FIELD ELECTROSTATICS */
1988             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1989
1990             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1991
1992             fscal            = felec;
1993
1994             fscal            = _mm_and_pd(fscal,cutoff_mask);
1995
1996             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1997
1998             /* Update vectorial force */
1999             fix2             = _mm_macc_pd(dx23,fscal,fix2);
2000             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
2001             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
2002
2003             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
2004             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
2005             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
2006
2007             }
2008
2009             /**************************
2010              * CALCULATE INTERACTIONS *
2011              **************************/
2012
2013             if (gmx_mm_any_lt(rsq31,rcutoff2))
2014             {
2015
2016             /* REACTION-FIELD ELECTROSTATICS */
2017             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
2018
2019             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
2020
2021             fscal            = felec;
2022
2023             fscal            = _mm_and_pd(fscal,cutoff_mask);
2024
2025             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2026
2027             /* Update vectorial force */
2028             fix3             = _mm_macc_pd(dx31,fscal,fix3);
2029             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
2030             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
2031
2032             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
2033             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
2034             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
2035
2036             }
2037
2038             /**************************
2039              * CALCULATE INTERACTIONS *
2040              **************************/
2041
2042             if (gmx_mm_any_lt(rsq32,rcutoff2))
2043             {
2044
2045             /* REACTION-FIELD ELECTROSTATICS */
2046             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
2047
2048             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
2049
2050             fscal            = felec;
2051
2052             fscal            = _mm_and_pd(fscal,cutoff_mask);
2053
2054             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2055
2056             /* Update vectorial force */
2057             fix3             = _mm_macc_pd(dx32,fscal,fix3);
2058             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
2059             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
2060
2061             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
2062             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
2063             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
2064
2065             }
2066
2067             /**************************
2068              * CALCULATE INTERACTIONS *
2069              **************************/
2070
2071             if (gmx_mm_any_lt(rsq33,rcutoff2))
2072             {
2073
2074             /* REACTION-FIELD ELECTROSTATICS */
2075             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
2076
2077             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
2078
2079             fscal            = felec;
2080
2081             fscal            = _mm_and_pd(fscal,cutoff_mask);
2082
2083             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2084
2085             /* Update vectorial force */
2086             fix3             = _mm_macc_pd(dx33,fscal,fix3);
2087             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
2088             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
2089
2090             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
2091             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
2092             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
2093
2094             }
2095
2096             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2097
2098             /* Inner loop uses 359 flops */
2099         }
2100
2101         /* End of innermost loop */
2102
2103         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2104                                               f+i_coord_offset,fshift+i_shift_offset);
2105
2106         /* Increment number of inner iterations */
2107         inneriter                  += j_index_end - j_index_start;
2108
2109         /* Outer loop uses 24 flops */
2110     }
2111
2112     /* Increment number of outer iterations */
2113     outeriter        += nri;
2114
2115     /* Update outer/inner flops */
2116
2117     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*359);
2118 }