src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_avx_128_fma_double.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_double kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_128_fma_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_VF_avx_128_fma_double
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Water4-Water4
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_VF_avx_128_fma_double
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB;
  74     int              j_coord_offsetA,j_coord_offsetB;
  75     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  76     real             rcutoff_scalar;
  77     real             *shiftvec,*fshift,*x,*f;
  78     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  79     int              vdwioffset0;
  80     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  81     int              vdwioffset1;
  82     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  83     int              vdwioffset2;
  84     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  85     int              vdwioffset3;
  86     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  87     int              vdwjidx0A,vdwjidx0B;
  88     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  89     int              vdwjidx1A,vdwjidx1B;
  90     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  91     int              vdwjidx2A,vdwjidx2B;
  92     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  93     int              vdwjidx3A,vdwjidx3B;
  94     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  95     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  96     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  97     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  98     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  99     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 100     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 101     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 102     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 103     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 104     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 105     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 106     real             *charge;
 107     int              nvdwtype;
 108     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 109     int              *vdwtype;
 110     real             *vdwparam;
 111     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 112     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 113     __m128d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 114     real             rswitch_scalar,d_scalar;
 115     __m128d          dummy_mask,cutoff_mask;
 116     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 117     __m128d          one     = _mm_set1_pd(1.0);
 118     __m128d          two     = _mm_set1_pd(2.0);
 119     x                = xx[0];
 120     f                = ff[0];
 121
 122     nri              = nlist->nri;
 123     iinr             = nlist->iinr;
 124     jindex           = nlist->jindex;
 125     jjnr             = nlist->jjnr;
 126     shiftidx         = nlist->shift;
 127     gid              = nlist->gid;
 128     shiftvec         = fr->shift_vec[0];
 129     fshift           = fr->fshift[0];
 130     facel            = _mm_set1_pd(fr->ic->epsfac);
 131     charge           = mdatoms->chargeA;
 132     krf              = _mm_set1_pd(fr->ic->k_rf);
 133     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 134     crf              = _mm_set1_pd(fr->ic->c_rf);
 135     nvdwtype         = fr->ntype;
 136     vdwparam         = fr->nbfp;
 137     vdwtype          = mdatoms->typeA;
 138
 139     /* Setup water-specific parameters */
 140     inr              = nlist->iinr[0];
 141     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 142     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 143     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 144     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 145
 146     jq1              = _mm_set1_pd(charge[inr+1]);
 147     jq2              = _mm_set1_pd(charge[inr+2]);
 148     jq3              = _mm_set1_pd(charge[inr+3]);
 149     vdwjidx0A        = 2*vdwtype[inr+0];
 150     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 151     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 152     qq11             = _mm_mul_pd(iq1,jq1);
 153     qq12             = _mm_mul_pd(iq1,jq2);
 154     qq13             = _mm_mul_pd(iq1,jq3);
 155     qq21             = _mm_mul_pd(iq2,jq1);
 156     qq22             = _mm_mul_pd(iq2,jq2);
 157     qq23             = _mm_mul_pd(iq2,jq3);
 158     qq31             = _mm_mul_pd(iq3,jq1);
 159     qq32             = _mm_mul_pd(iq3,jq2);
 160     qq33             = _mm_mul_pd(iq3,jq3);
 161
 162     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 163     rcutoff_scalar   = fr->ic->rcoulomb;
 164     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 165     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 166
 167     rswitch_scalar   = fr->ic->rvdw_switch;
 168     rswitch          = _mm_set1_pd(rswitch_scalar);
 169     /* Setup switch parameters */
 170     d_scalar         = rcutoff_scalar-rswitch_scalar;
 171     d                = _mm_set1_pd(d_scalar);
 172     swV3             = _mm_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 173     swV4             = _mm_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 174     swV5             = _mm_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 175     swF2             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 176     swF3             = _mm_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 177     swF4             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 178
 179     /* Avoid stupid compiler warnings */
 180     jnrA = jnrB = 0;
 181     j_coord_offsetA = 0;
 182     j_coord_offsetB = 0;
 183
 184     outeriter        = 0;
 185     inneriter        = 0;
 186
 187     /* Start outer loop over neighborlists */
 188     for(iidx=0; iidx<nri; iidx++)
 189     {
 190         /* Load shift vector for this list */
 191         i_shift_offset   = DIM*shiftidx[iidx];
 192
 193         /* Load limits for loop over neighbors */
 194         j_index_start    = jindex[iidx];
 195         j_index_end      = jindex[iidx+1];
 196
 197         /* Get outer coordinate index */
 198         inr              = iinr[iidx];
 199         i_coord_offset   = DIM*inr;
 200
 201         /* Load i particle coords and add shift vector */
 202         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 203                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 204
 205         fix0             = _mm_setzero_pd();
 206         fiy0             = _mm_setzero_pd();
 207         fiz0             = _mm_setzero_pd();
 208         fix1             = _mm_setzero_pd();
 209         fiy1             = _mm_setzero_pd();
 210         fiz1             = _mm_setzero_pd();
 211         fix2             = _mm_setzero_pd();
 212         fiy2             = _mm_setzero_pd();
 213         fiz2             = _mm_setzero_pd();
 214         fix3             = _mm_setzero_pd();
 215         fiy3             = _mm_setzero_pd();
 216         fiz3             = _mm_setzero_pd();
 217
 218         /* Reset potential sums */
 219         velecsum         = _mm_setzero_pd();
 220         vvdwsum          = _mm_setzero_pd();
 221
 222         /* Start inner kernel loop */
 223         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 224         {
 225
 226             /* Get j neighbor index, and coordinate index */
 227             jnrA             = jjnr[jidx];
 228             jnrB             = jjnr[jidx+1];
 229             j_coord_offsetA  = DIM*jnrA;
 230             j_coord_offsetB  = DIM*jnrB;
 231
 232             /* load j atom coordinates */
 233             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 234                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 235                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 236
 237             /* Calculate displacement vector */
 238             dx00             = _mm_sub_pd(ix0,jx0);
 239             dy00             = _mm_sub_pd(iy0,jy0);
 240             dz00             = _mm_sub_pd(iz0,jz0);
 241             dx11             = _mm_sub_pd(ix1,jx1);
 242             dy11             = _mm_sub_pd(iy1,jy1);
 243             dz11             = _mm_sub_pd(iz1,jz1);
 244             dx12             = _mm_sub_pd(ix1,jx2);
 245             dy12             = _mm_sub_pd(iy1,jy2);
 246             dz12             = _mm_sub_pd(iz1,jz2);
 247             dx13             = _mm_sub_pd(ix1,jx3);
 248             dy13             = _mm_sub_pd(iy1,jy3);
 249             dz13             = _mm_sub_pd(iz1,jz3);
 250             dx21             = _mm_sub_pd(ix2,jx1);
 251             dy21             = _mm_sub_pd(iy2,jy1);
 252             dz21             = _mm_sub_pd(iz2,jz1);
 253             dx22             = _mm_sub_pd(ix2,jx2);
 254             dy22             = _mm_sub_pd(iy2,jy2);
 255             dz22             = _mm_sub_pd(iz2,jz2);
 256             dx23             = _mm_sub_pd(ix2,jx3);
 257             dy23             = _mm_sub_pd(iy2,jy3);
 258             dz23             = _mm_sub_pd(iz2,jz3);
 259             dx31             = _mm_sub_pd(ix3,jx1);
 260             dy31             = _mm_sub_pd(iy3,jy1);
 261             dz31             = _mm_sub_pd(iz3,jz1);
 262             dx32             = _mm_sub_pd(ix3,jx2);
 263             dy32             = _mm_sub_pd(iy3,jy2);
 264             dz32             = _mm_sub_pd(iz3,jz2);
 265             dx33             = _mm_sub_pd(ix3,jx3);
 266             dy33             = _mm_sub_pd(iy3,jy3);
 267             dz33             = _mm_sub_pd(iz3,jz3);
 268
 269             /* Calculate squared distance and things based on it */
 270             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 271             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 272             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 273             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 274             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 275             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 276             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 277             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 278             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 279             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 280
 281             rinv00           = avx128fma_invsqrt_d(rsq00);
 282             rinv11           = avx128fma_invsqrt_d(rsq11);
 283             rinv12           = avx128fma_invsqrt_d(rsq12);
 284             rinv13           = avx128fma_invsqrt_d(rsq13);
 285             rinv21           = avx128fma_invsqrt_d(rsq21);
 286             rinv22           = avx128fma_invsqrt_d(rsq22);
 287             rinv23           = avx128fma_invsqrt_d(rsq23);
 288             rinv31           = avx128fma_invsqrt_d(rsq31);
 289             rinv32           = avx128fma_invsqrt_d(rsq32);
 290             rinv33           = avx128fma_invsqrt_d(rsq33);
 291
 292             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 293             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 294             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 295             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 296             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 297             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 298             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 299             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 300             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 301             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 302
 303             fjx0             = _mm_setzero_pd();
 304             fjy0             = _mm_setzero_pd();
 305             fjz0             = _mm_setzero_pd();
 306             fjx1             = _mm_setzero_pd();
 307             fjy1             = _mm_setzero_pd();
 308             fjz1             = _mm_setzero_pd();
 309             fjx2             = _mm_setzero_pd();
 310             fjy2             = _mm_setzero_pd();
 311             fjz2             = _mm_setzero_pd();
 312             fjx3             = _mm_setzero_pd();
 313             fjy3             = _mm_setzero_pd();
 314             fjz3             = _mm_setzero_pd();
 315
 316             /**************************
 317              * CALCULATE INTERACTIONS *
 318              **************************/
 319
 320             if (gmx_mm_any_lt(rsq00,rcutoff2))
 321             {
 322
 323             r00              = _mm_mul_pd(rsq00,rinv00);
 324
 325             /* LENNARD-JONES DISPERSION/REPULSION */
 326
 327             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 328             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 329             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 330             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 331             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 332
 333             d                = _mm_sub_pd(r00,rswitch);
 334             d                = _mm_max_pd(d,_mm_setzero_pd());
 335             d2               = _mm_mul_pd(d,d);
 336             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 337
 338             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 339
 340             /* Evaluate switch function */
 341             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 342             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 343             vvdw             = _mm_mul_pd(vvdw,sw);
 344             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 345
 346             /* Update potential sum for this i atom from the interaction with this j atom. */
 347             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 348             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 349
 350             fscal            = fvdw;
 351
 352             fscal            = _mm_and_pd(fscal,cutoff_mask);
 353
 354             /* Update vectorial force */
 355             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 356             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 357             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 358
 359             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 360             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 361             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 362
 363             }
 364
 365             /**************************
 366              * CALCULATE INTERACTIONS *
 367              **************************/
 368
 369             if (gmx_mm_any_lt(rsq11,rcutoff2))
 370             {
 371
 372             /* REACTION-FIELD ELECTROSTATICS */
 373             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 374             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 375
 376             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 377
 378             /* Update potential sum for this i atom from the interaction with this j atom. */
 379             velec            = _mm_and_pd(velec,cutoff_mask);
 380             velecsum         = _mm_add_pd(velecsum,velec);
 381
 382             fscal            = felec;
 383
 384             fscal            = _mm_and_pd(fscal,cutoff_mask);
 385
 386             /* Update vectorial force */
 387             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 388             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 389             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 390
 391             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 392             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 393             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 394
 395             }
 396
 397             /**************************
 398              * CALCULATE INTERACTIONS *
 399              **************************/
 400
 401             if (gmx_mm_any_lt(rsq12,rcutoff2))
 402             {
 403
 404             /* REACTION-FIELD ELECTROSTATICS */
 405             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 406             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 407
 408             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 409
 410             /* Update potential sum for this i atom from the interaction with this j atom. */
 411             velec            = _mm_and_pd(velec,cutoff_mask);
 412             velecsum         = _mm_add_pd(velecsum,velec);
 413
 414             fscal            = felec;
 415
 416             fscal            = _mm_and_pd(fscal,cutoff_mask);
 417
 418             /* Update vectorial force */
 419             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 420             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 421             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 422
 423             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 424             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 425             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 426
 427             }
 428
 429             /**************************
 430              * CALCULATE INTERACTIONS *
 431              **************************/
 432
 433             if (gmx_mm_any_lt(rsq13,rcutoff2))
 434             {
 435
 436             /* REACTION-FIELD ELECTROSTATICS */
 437             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 438             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 439
 440             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 441
 442             /* Update potential sum for this i atom from the interaction with this j atom. */
 443             velec            = _mm_and_pd(velec,cutoff_mask);
 444             velecsum         = _mm_add_pd(velecsum,velec);
 445
 446             fscal            = felec;
 447
 448             fscal            = _mm_and_pd(fscal,cutoff_mask);
 449
 450             /* Update vectorial force */
 451             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 452             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 453             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 454
 455             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 456             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 457             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 458
 459             }
 460
 461             /**************************
 462              * CALCULATE INTERACTIONS *
 463              **************************/
 464
 465             if (gmx_mm_any_lt(rsq21,rcutoff2))
 466             {
 467
 468             /* REACTION-FIELD ELECTROSTATICS */
 469             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 470             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 471
 472             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 473
 474             /* Update potential sum for this i atom from the interaction with this j atom. */
 475             velec            = _mm_and_pd(velec,cutoff_mask);
 476             velecsum         = _mm_add_pd(velecsum,velec);
 477
 478             fscal            = felec;
 479
 480             fscal            = _mm_and_pd(fscal,cutoff_mask);
 481
 482             /* Update vectorial force */
 483             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 484             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 485             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 486
 487             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 488             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 489             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 490
 491             }
 492
 493             /**************************
 494              * CALCULATE INTERACTIONS *
 495              **************************/
 496
 497             if (gmx_mm_any_lt(rsq22,rcutoff2))
 498             {
 499
 500             /* REACTION-FIELD ELECTROSTATICS */
 501             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 502             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 503
 504             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 505
 506             /* Update potential sum for this i atom from the interaction with this j atom. */
 507             velec            = _mm_and_pd(velec,cutoff_mask);
 508             velecsum         = _mm_add_pd(velecsum,velec);
 509
 510             fscal            = felec;
 511
 512             fscal            = _mm_and_pd(fscal,cutoff_mask);
 513
 514             /* Update vectorial force */
 515             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 516             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 517             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 518
 519             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 520             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 521             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 522
 523             }
 524
 525             /**************************
 526              * CALCULATE INTERACTIONS *
 527              **************************/
 528
 529             if (gmx_mm_any_lt(rsq23,rcutoff2))
 530             {
 531
 532             /* REACTION-FIELD ELECTROSTATICS */
 533             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 534             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 535
 536             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 537
 538             /* Update potential sum for this i atom from the interaction with this j atom. */
 539             velec            = _mm_and_pd(velec,cutoff_mask);
 540             velecsum         = _mm_add_pd(velecsum,velec);
 541
 542             fscal            = felec;
 543
 544             fscal            = _mm_and_pd(fscal,cutoff_mask);
 545
 546             /* Update vectorial force */
 547             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 548             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 549             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 550
 551             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 552             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 553             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 554
 555             }
 556
 557             /**************************
 558              * CALCULATE INTERACTIONS *
 559              **************************/
 560
 561             if (gmx_mm_any_lt(rsq31,rcutoff2))
 562             {
 563
 564             /* REACTION-FIELD ELECTROSTATICS */
 565             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
 566             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
 567
 568             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
 569
 570             /* Update potential sum for this i atom from the interaction with this j atom. */
 571             velec            = _mm_and_pd(velec,cutoff_mask);
 572             velecsum         = _mm_add_pd(velecsum,velec);
 573
 574             fscal            = felec;
 575
 576             fscal            = _mm_and_pd(fscal,cutoff_mask);
 577
 578             /* Update vectorial force */
 579             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 580             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 581             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 582
 583             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 584             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 585             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 586
 587             }
 588
 589             /**************************
 590              * CALCULATE INTERACTIONS *
 591              **************************/
 592
 593             if (gmx_mm_any_lt(rsq32,rcutoff2))
 594             {
 595
 596             /* REACTION-FIELD ELECTROSTATICS */
 597             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
 598             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
 599
 600             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
 601
 602             /* Update potential sum for this i atom from the interaction with this j atom. */
 603             velec            = _mm_and_pd(velec,cutoff_mask);
 604             velecsum         = _mm_add_pd(velecsum,velec);
 605
 606             fscal            = felec;
 607
 608             fscal            = _mm_and_pd(fscal,cutoff_mask);
 609
 610             /* Update vectorial force */
 611             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 612             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 613             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 614
 615             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 616             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 617             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 618
 619             }
 620
 621             /**************************
 622              * CALCULATE INTERACTIONS *
 623              **************************/
 624
 625             if (gmx_mm_any_lt(rsq33,rcutoff2))
 626             {
 627
 628             /* REACTION-FIELD ELECTROSTATICS */
 629             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
 630             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
 631
 632             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
 633
 634             /* Update potential sum for this i atom from the interaction with this j atom. */
 635             velec            = _mm_and_pd(velec,cutoff_mask);
 636             velecsum         = _mm_add_pd(velecsum,velec);
 637
 638             fscal            = felec;
 639
 640             fscal            = _mm_and_pd(fscal,cutoff_mask);
 641
 642             /* Update vectorial force */
 643             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 644             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 645             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 646
 647             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 648             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 649             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 650
 651             }
 652
 653             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);
 654
 655             /* Inner loop uses 416 flops */
 656         }
 657
 658         if(jidx<j_index_end)
 659         {
 660
 661             jnrA             = jjnr[jidx];
 662             j_coord_offsetA  = DIM*jnrA;
 663
 664             /* load j atom coordinates */
 665             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 666                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 667                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 668
 669             /* Calculate displacement vector */
 670             dx00             = _mm_sub_pd(ix0,jx0);
 671             dy00             = _mm_sub_pd(iy0,jy0);
 672             dz00             = _mm_sub_pd(iz0,jz0);
 673             dx11             = _mm_sub_pd(ix1,jx1);
 674             dy11             = _mm_sub_pd(iy1,jy1);
 675             dz11             = _mm_sub_pd(iz1,jz1);
 676             dx12             = _mm_sub_pd(ix1,jx2);
 677             dy12             = _mm_sub_pd(iy1,jy2);
 678             dz12             = _mm_sub_pd(iz1,jz2);
 679             dx13             = _mm_sub_pd(ix1,jx3);
 680             dy13             = _mm_sub_pd(iy1,jy3);
 681             dz13             = _mm_sub_pd(iz1,jz3);
 682             dx21             = _mm_sub_pd(ix2,jx1);
 683             dy21             = _mm_sub_pd(iy2,jy1);
 684             dz21             = _mm_sub_pd(iz2,jz1);
 685             dx22             = _mm_sub_pd(ix2,jx2);
 686             dy22             = _mm_sub_pd(iy2,jy2);
 687             dz22             = _mm_sub_pd(iz2,jz2);
 688             dx23             = _mm_sub_pd(ix2,jx3);
 689             dy23             = _mm_sub_pd(iy2,jy3);
 690             dz23             = _mm_sub_pd(iz2,jz3);
 691             dx31             = _mm_sub_pd(ix3,jx1);
 692             dy31             = _mm_sub_pd(iy3,jy1);
 693             dz31             = _mm_sub_pd(iz3,jz1);
 694             dx32             = _mm_sub_pd(ix3,jx2);
 695             dy32             = _mm_sub_pd(iy3,jy2);
 696             dz32             = _mm_sub_pd(iz3,jz2);
 697             dx33             = _mm_sub_pd(ix3,jx3);
 698             dy33             = _mm_sub_pd(iy3,jy3);
 699             dz33             = _mm_sub_pd(iz3,jz3);
 700
 701             /* Calculate squared distance and things based on it */
 702             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 703             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 704             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 705             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 706             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 707             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 708             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 709             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 710             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 711             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 712
 713             rinv00           = avx128fma_invsqrt_d(rsq00);
 714             rinv11           = avx128fma_invsqrt_d(rsq11);
 715             rinv12           = avx128fma_invsqrt_d(rsq12);
 716             rinv13           = avx128fma_invsqrt_d(rsq13);
 717             rinv21           = avx128fma_invsqrt_d(rsq21);
 718             rinv22           = avx128fma_invsqrt_d(rsq22);
 719             rinv23           = avx128fma_invsqrt_d(rsq23);
 720             rinv31           = avx128fma_invsqrt_d(rsq31);
 721             rinv32           = avx128fma_invsqrt_d(rsq32);
 722             rinv33           = avx128fma_invsqrt_d(rsq33);
 723
 724             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 725             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 726             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 727             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 728             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 729             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 730             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 731             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 732             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 733             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 734
 735             fjx0             = _mm_setzero_pd();
 736             fjy0             = _mm_setzero_pd();
 737             fjz0             = _mm_setzero_pd();
 738             fjx1             = _mm_setzero_pd();
 739             fjy1             = _mm_setzero_pd();
 740             fjz1             = _mm_setzero_pd();
 741             fjx2             = _mm_setzero_pd();
 742             fjy2             = _mm_setzero_pd();
 743             fjz2             = _mm_setzero_pd();
 744             fjx3             = _mm_setzero_pd();
 745             fjy3             = _mm_setzero_pd();
 746             fjz3             = _mm_setzero_pd();
 747
 748             /**************************
 749              * CALCULATE INTERACTIONS *
 750              **************************/
 751
 752             if (gmx_mm_any_lt(rsq00,rcutoff2))
 753             {
 754
 755             r00              = _mm_mul_pd(rsq00,rinv00);
 756
 757             /* LENNARD-JONES DISPERSION/REPULSION */
 758
 759             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 760             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 761             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 762             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 763             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 764
 765             d                = _mm_sub_pd(r00,rswitch);
 766             d                = _mm_max_pd(d,_mm_setzero_pd());
 767             d2               = _mm_mul_pd(d,d);
 768             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 769
 770             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 771
 772             /* Evaluate switch function */
 773             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 774             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 775             vvdw             = _mm_mul_pd(vvdw,sw);
 776             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 777
 778             /* Update potential sum for this i atom from the interaction with this j atom. */
 779             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 780             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 781             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 782
 783             fscal            = fvdw;
 784
 785             fscal            = _mm_and_pd(fscal,cutoff_mask);
 786
 787             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 788
 789             /* Update vectorial force */
 790             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 791             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 792             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 793
 794             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 795             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 796             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 797
 798             }
 799
 800             /**************************
 801              * CALCULATE INTERACTIONS *
 802              **************************/
 803
 804             if (gmx_mm_any_lt(rsq11,rcutoff2))
 805             {
 806
 807             /* REACTION-FIELD ELECTROSTATICS */
 808             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 809             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 810
 811             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 812
 813             /* Update potential sum for this i atom from the interaction with this j atom. */
 814             velec            = _mm_and_pd(velec,cutoff_mask);
 815             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 816             velecsum         = _mm_add_pd(velecsum,velec);
 817
 818             fscal            = felec;
 819
 820             fscal            = _mm_and_pd(fscal,cutoff_mask);
 821
 822             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 823
 824             /* Update vectorial force */
 825             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 826             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 827             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 828
 829             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 830             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 831             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 832
 833             }
 834
 835             /**************************
 836              * CALCULATE INTERACTIONS *
 837              **************************/
 838
 839             if (gmx_mm_any_lt(rsq12,rcutoff2))
 840             {
 841
 842             /* REACTION-FIELD ELECTROSTATICS */
 843             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 844             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 845
 846             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 847
 848             /* Update potential sum for this i atom from the interaction with this j atom. */
 849             velec            = _mm_and_pd(velec,cutoff_mask);
 850             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 851             velecsum         = _mm_add_pd(velecsum,velec);
 852
 853             fscal            = felec;
 854
 855             fscal            = _mm_and_pd(fscal,cutoff_mask);
 856
 857             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 858
 859             /* Update vectorial force */
 860             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 861             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 862             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 863
 864             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 865             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 866             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 867
 868             }
 869
 870             /**************************
 871              * CALCULATE INTERACTIONS *
 872              **************************/
 873
 874             if (gmx_mm_any_lt(rsq13,rcutoff2))
 875             {
 876
 877             /* REACTION-FIELD ELECTROSTATICS */
 878             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 879             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 880
 881             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 882
 883             /* Update potential sum for this i atom from the interaction with this j atom. */
 884             velec            = _mm_and_pd(velec,cutoff_mask);
 885             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 886             velecsum         = _mm_add_pd(velecsum,velec);
 887
 888             fscal            = felec;
 889
 890             fscal            = _mm_and_pd(fscal,cutoff_mask);
 891
 892             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 893
 894             /* Update vectorial force */
 895             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 896             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 897             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 898
 899             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 900             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 901             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 902
 903             }
 904
 905             /**************************
 906              * CALCULATE INTERACTIONS *
 907              **************************/
 908
 909             if (gmx_mm_any_lt(rsq21,rcutoff2))
 910             {
 911
 912             /* REACTION-FIELD ELECTROSTATICS */
 913             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 914             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 915
 916             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 917
 918             /* Update potential sum for this i atom from the interaction with this j atom. */
 919             velec            = _mm_and_pd(velec,cutoff_mask);
 920             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 921             velecsum         = _mm_add_pd(velecsum,velec);
 922
 923             fscal            = felec;
 924
 925             fscal            = _mm_and_pd(fscal,cutoff_mask);
 926
 927             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 928
 929             /* Update vectorial force */
 930             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 931             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 932             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 933
 934             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 935             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 936             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 937
 938             }
 939
 940             /**************************
 941              * CALCULATE INTERACTIONS *
 942              **************************/
 943
 944             if (gmx_mm_any_lt(rsq22,rcutoff2))
 945             {
 946
 947             /* REACTION-FIELD ELECTROSTATICS */
 948             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 949             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 950
 951             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 952
 953             /* Update potential sum for this i atom from the interaction with this j atom. */
 954             velec            = _mm_and_pd(velec,cutoff_mask);
 955             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 956             velecsum         = _mm_add_pd(velecsum,velec);
 957
 958             fscal            = felec;
 959
 960             fscal            = _mm_and_pd(fscal,cutoff_mask);
 961
 962             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 963
 964             /* Update vectorial force */
 965             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 966             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 967             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 968
 969             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 970             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 971             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 972
 973             }
 974
 975             /**************************
 976              * CALCULATE INTERACTIONS *
 977              **************************/
 978
 979             if (gmx_mm_any_lt(rsq23,rcutoff2))
 980             {
 981
 982             /* REACTION-FIELD ELECTROSTATICS */
 983             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 984             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 985
 986             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 987
 988             /* Update potential sum for this i atom from the interaction with this j atom. */
 989             velec            = _mm_and_pd(velec,cutoff_mask);
 990             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 991             velecsum         = _mm_add_pd(velecsum,velec);
 992
 993             fscal            = felec;
 994
 995             fscal            = _mm_and_pd(fscal,cutoff_mask);
 996
 997             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 998
 999             /* Update vectorial force */
1000             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1001             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1002             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1003
1004             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1005             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1006             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1007
1008             }
1009
1010             /**************************
1011              * CALCULATE INTERACTIONS *
1012              **************************/
1013
1014             if (gmx_mm_any_lt(rsq31,rcutoff2))
1015             {
1016
1017             /* REACTION-FIELD ELECTROSTATICS */
1018             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
1019             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1020
1021             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1022
1023             /* Update potential sum for this i atom from the interaction with this j atom. */
1024             velec            = _mm_and_pd(velec,cutoff_mask);
1025             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1026             velecsum         = _mm_add_pd(velecsum,velec);
1027
1028             fscal            = felec;
1029
1030             fscal            = _mm_and_pd(fscal,cutoff_mask);
1031
1032             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1033
1034             /* Update vectorial force */
1035             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1036             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1037             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1038
1039             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1040             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1041             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1042
1043             }
1044
1045             /**************************
1046              * CALCULATE INTERACTIONS *
1047              **************************/
1048
1049             if (gmx_mm_any_lt(rsq32,rcutoff2))
1050             {
1051
1052             /* REACTION-FIELD ELECTROSTATICS */
1053             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
1054             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1055
1056             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1057
1058             /* Update potential sum for this i atom from the interaction with this j atom. */
1059             velec            = _mm_and_pd(velec,cutoff_mask);
1060             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1061             velecsum         = _mm_add_pd(velecsum,velec);
1062
1063             fscal            = felec;
1064
1065             fscal            = _mm_and_pd(fscal,cutoff_mask);
1066
1067             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1068
1069             /* Update vectorial force */
1070             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1071             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1072             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1073
1074             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1075             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1076             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1077
1078             }
1079
1080             /**************************
1081              * CALCULATE INTERACTIONS *
1082              **************************/
1083
1084             if (gmx_mm_any_lt(rsq33,rcutoff2))
1085             {
1086
1087             /* REACTION-FIELD ELECTROSTATICS */
1088             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
1089             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1090
1091             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1092
1093             /* Update potential sum for this i atom from the interaction with this j atom. */
1094             velec            = _mm_and_pd(velec,cutoff_mask);
1095             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1096             velecsum         = _mm_add_pd(velecsum,velec);
1097
1098             fscal            = felec;
1099
1100             fscal            = _mm_and_pd(fscal,cutoff_mask);
1101
1102             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1103
1104             /* Update vectorial force */
1105             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1106             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1107             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1108
1109             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1110             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1111             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1112
1113             }
1114
1115             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1116
1117             /* Inner loop uses 416 flops */
1118         }
1119
1120         /* End of innermost loop */
1121
1122         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1123                                               f+i_coord_offset,fshift+i_shift_offset);
1124
1125         ggid                        = gid[iidx];
1126         /* Update potential energies */
1127         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1128         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
1129
1130         /* Increment number of inner iterations */
1131         inneriter                  += j_index_end - j_index_start;
1132
1133         /* Outer loop uses 26 flops */
1134     }
1135
1136     /* Increment number of outer iterations */
1137     outeriter        += nri;
1138
1139     /* Update outer/inner flops */
1140
1141     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*416);
1142 }
1143 /*
1144  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_F_avx_128_fma_double
1145  * Electrostatics interaction: ReactionField
1146  * VdW interaction:            LennardJones
1147  * Geometry:                   Water4-Water4
1148  * Calculate force/pot:        Force
1149  */
1150 void
1151 nb_kernel_ElecRFCut_VdwLJSw_GeomW4W4_F_avx_128_fma_double
1152                     (t_nblist                    * gmx_restrict       nlist,
1153                      rvec                        * gmx_restrict          xx,
1154                      rvec                        * gmx_restrict          ff,
1155                      struct t_forcerec           * gmx_restrict          fr,
1156                      t_mdatoms                   * gmx_restrict     mdatoms,
1157                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1158                      t_nrnb                      * gmx_restrict        nrnb)
1159 {
1160     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1161      * just 0 for non-waters.
1162      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1163      * jnr indices corresponding to data put in the four positions in the SIMD register.
1164      */
1165     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1166     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1167     int              jnrA,jnrB;
1168     int              j_coord_offsetA,j_coord_offsetB;
1169     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1170     real             rcutoff_scalar;
1171     real             *shiftvec,*fshift,*x,*f;
1172     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1173     int              vdwioffset0;
1174     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1175     int              vdwioffset1;
1176     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1177     int              vdwioffset2;
1178     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1179     int              vdwioffset3;
1180     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1181     int              vdwjidx0A,vdwjidx0B;
1182     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1183     int              vdwjidx1A,vdwjidx1B;
1184     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1185     int              vdwjidx2A,vdwjidx2B;
1186     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1187     int              vdwjidx3A,vdwjidx3B;
1188     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1189     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1190     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1191     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1192     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1193     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1194     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1195     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1196     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1197     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1198     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1199     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1200     real             *charge;
1201     int              nvdwtype;
1202     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1203     int              *vdwtype;
1204     real             *vdwparam;
1205     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1206     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1207     __m128d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
1208     real             rswitch_scalar,d_scalar;
1209     __m128d          dummy_mask,cutoff_mask;
1210     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1211     __m128d          one     = _mm_set1_pd(1.0);
1212     __m128d          two     = _mm_set1_pd(2.0);
1213     x                = xx[0];
1214     f                = ff[0];
1215
1216     nri              = nlist->nri;
1217     iinr             = nlist->iinr;
1218     jindex           = nlist->jindex;
1219     jjnr             = nlist->jjnr;
1220     shiftidx         = nlist->shift;
1221     gid              = nlist->gid;
1222     shiftvec         = fr->shift_vec[0];
1223     fshift           = fr->fshift[0];
1224     facel            = _mm_set1_pd(fr->ic->epsfac);
1225     charge           = mdatoms->chargeA;
1226     krf              = _mm_set1_pd(fr->ic->k_rf);
1227     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1228     crf              = _mm_set1_pd(fr->ic->c_rf);
1229     nvdwtype         = fr->ntype;
1230     vdwparam         = fr->nbfp;
1231     vdwtype          = mdatoms->typeA;
1232
1233     /* Setup water-specific parameters */
1234     inr              = nlist->iinr[0];
1235     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1236     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1237     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1238     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1239
1240     jq1              = _mm_set1_pd(charge[inr+1]);
1241     jq2              = _mm_set1_pd(charge[inr+2]);
1242     jq3              = _mm_set1_pd(charge[inr+3]);
1243     vdwjidx0A        = 2*vdwtype[inr+0];
1244     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1245     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1246     qq11             = _mm_mul_pd(iq1,jq1);
1247     qq12             = _mm_mul_pd(iq1,jq2);
1248     qq13             = _mm_mul_pd(iq1,jq3);
1249     qq21             = _mm_mul_pd(iq2,jq1);
1250     qq22             = _mm_mul_pd(iq2,jq2);
1251     qq23             = _mm_mul_pd(iq2,jq3);
1252     qq31             = _mm_mul_pd(iq3,jq1);
1253     qq32             = _mm_mul_pd(iq3,jq2);
1254     qq33             = _mm_mul_pd(iq3,jq3);
1255
1256     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1257     rcutoff_scalar   = fr->ic->rcoulomb;
1258     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1259     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1260
1261     rswitch_scalar   = fr->ic->rvdw_switch;
1262     rswitch          = _mm_set1_pd(rswitch_scalar);
1263     /* Setup switch parameters */
1264     d_scalar         = rcutoff_scalar-rswitch_scalar;
1265     d                = _mm_set1_pd(d_scalar);
1266     swV3             = _mm_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
1267     swV4             = _mm_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1268     swV5             = _mm_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1269     swF2             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
1270     swF3             = _mm_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1271     swF4             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1272
1273     /* Avoid stupid compiler warnings */
1274     jnrA = jnrB = 0;
1275     j_coord_offsetA = 0;
1276     j_coord_offsetB = 0;
1277
1278     outeriter        = 0;
1279     inneriter        = 0;
1280
1281     /* Start outer loop over neighborlists */
1282     for(iidx=0; iidx<nri; iidx++)
1283     {
1284         /* Load shift vector for this list */
1285         i_shift_offset   = DIM*shiftidx[iidx];
1286
1287         /* Load limits for loop over neighbors */
1288         j_index_start    = jindex[iidx];
1289         j_index_end      = jindex[iidx+1];
1290
1291         /* Get outer coordinate index */
1292         inr              = iinr[iidx];
1293         i_coord_offset   = DIM*inr;
1294
1295         /* Load i particle coords and add shift vector */
1296         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1297                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1298
1299         fix0             = _mm_setzero_pd();
1300         fiy0             = _mm_setzero_pd();
1301         fiz0             = _mm_setzero_pd();
1302         fix1             = _mm_setzero_pd();
1303         fiy1             = _mm_setzero_pd();
1304         fiz1             = _mm_setzero_pd();
1305         fix2             = _mm_setzero_pd();
1306         fiy2             = _mm_setzero_pd();
1307         fiz2             = _mm_setzero_pd();
1308         fix3             = _mm_setzero_pd();
1309         fiy3             = _mm_setzero_pd();
1310         fiz3             = _mm_setzero_pd();
1311
1312         /* Start inner kernel loop */
1313         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1314         {
1315
1316             /* Get j neighbor index, and coordinate index */
1317             jnrA             = jjnr[jidx];
1318             jnrB             = jjnr[jidx+1];
1319             j_coord_offsetA  = DIM*jnrA;
1320             j_coord_offsetB  = DIM*jnrB;
1321
1322             /* load j atom coordinates */
1323             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1324                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1325                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1326
1327             /* Calculate displacement vector */
1328             dx00             = _mm_sub_pd(ix0,jx0);
1329             dy00             = _mm_sub_pd(iy0,jy0);
1330             dz00             = _mm_sub_pd(iz0,jz0);
1331             dx11             = _mm_sub_pd(ix1,jx1);
1332             dy11             = _mm_sub_pd(iy1,jy1);
1333             dz11             = _mm_sub_pd(iz1,jz1);
1334             dx12             = _mm_sub_pd(ix1,jx2);
1335             dy12             = _mm_sub_pd(iy1,jy2);
1336             dz12             = _mm_sub_pd(iz1,jz2);
1337             dx13             = _mm_sub_pd(ix1,jx3);
1338             dy13             = _mm_sub_pd(iy1,jy3);
1339             dz13             = _mm_sub_pd(iz1,jz3);
1340             dx21             = _mm_sub_pd(ix2,jx1);
1341             dy21             = _mm_sub_pd(iy2,jy1);
1342             dz21             = _mm_sub_pd(iz2,jz1);
1343             dx22             = _mm_sub_pd(ix2,jx2);
1344             dy22             = _mm_sub_pd(iy2,jy2);
1345             dz22             = _mm_sub_pd(iz2,jz2);
1346             dx23             = _mm_sub_pd(ix2,jx3);
1347             dy23             = _mm_sub_pd(iy2,jy3);
1348             dz23             = _mm_sub_pd(iz2,jz3);
1349             dx31             = _mm_sub_pd(ix3,jx1);
1350             dy31             = _mm_sub_pd(iy3,jy1);
1351             dz31             = _mm_sub_pd(iz3,jz1);
1352             dx32             = _mm_sub_pd(ix3,jx2);
1353             dy32             = _mm_sub_pd(iy3,jy2);
1354             dz32             = _mm_sub_pd(iz3,jz2);
1355             dx33             = _mm_sub_pd(ix3,jx3);
1356             dy33             = _mm_sub_pd(iy3,jy3);
1357             dz33             = _mm_sub_pd(iz3,jz3);
1358
1359             /* Calculate squared distance and things based on it */
1360             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1361             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1362             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1363             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1364             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1365             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1366             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1367             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1368             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1369             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1370
1371             rinv00           = avx128fma_invsqrt_d(rsq00);
1372             rinv11           = avx128fma_invsqrt_d(rsq11);
1373             rinv12           = avx128fma_invsqrt_d(rsq12);
1374             rinv13           = avx128fma_invsqrt_d(rsq13);
1375             rinv21           = avx128fma_invsqrt_d(rsq21);
1376             rinv22           = avx128fma_invsqrt_d(rsq22);
1377             rinv23           = avx128fma_invsqrt_d(rsq23);
1378             rinv31           = avx128fma_invsqrt_d(rsq31);
1379             rinv32           = avx128fma_invsqrt_d(rsq32);
1380             rinv33           = avx128fma_invsqrt_d(rsq33);
1381
1382             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1383             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1384             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1385             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1386             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1387             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1388             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1389             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1390             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1391             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1392
1393             fjx0             = _mm_setzero_pd();
1394             fjy0             = _mm_setzero_pd();
1395             fjz0             = _mm_setzero_pd();
1396             fjx1             = _mm_setzero_pd();
1397             fjy1             = _mm_setzero_pd();
1398             fjz1             = _mm_setzero_pd();
1399             fjx2             = _mm_setzero_pd();
1400             fjy2             = _mm_setzero_pd();
1401             fjz2             = _mm_setzero_pd();
1402             fjx3             = _mm_setzero_pd();
1403             fjy3             = _mm_setzero_pd();
1404             fjz3             = _mm_setzero_pd();
1405
1406             /**************************
1407              * CALCULATE INTERACTIONS *
1408              **************************/
1409
1410             if (gmx_mm_any_lt(rsq00,rcutoff2))
1411             {
1412
1413             r00              = _mm_mul_pd(rsq00,rinv00);
1414
1415             /* LENNARD-JONES DISPERSION/REPULSION */
1416
1417             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1418             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
1419             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
1420             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
1421             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
1422
1423             d                = _mm_sub_pd(r00,rswitch);
1424             d                = _mm_max_pd(d,_mm_setzero_pd());
1425             d2               = _mm_mul_pd(d,d);
1426             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
1427
1428             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
1429
1430             /* Evaluate switch function */
1431             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1432             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
1433             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1434
1435             fscal            = fvdw;
1436
1437             fscal            = _mm_and_pd(fscal,cutoff_mask);
1438
1439             /* Update vectorial force */
1440             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1441             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1442             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1443
1444             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1445             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1446             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1447
1448             }
1449
1450             /**************************
1451              * CALCULATE INTERACTIONS *
1452              **************************/
1453
1454             if (gmx_mm_any_lt(rsq11,rcutoff2))
1455             {
1456
1457             /* REACTION-FIELD ELECTROSTATICS */
1458             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1459
1460             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1461
1462             fscal            = felec;
1463
1464             fscal            = _mm_and_pd(fscal,cutoff_mask);
1465
1466             /* Update vectorial force */
1467             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1468             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1469             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1470
1471             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1472             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1473             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1474
1475             }
1476
1477             /**************************
1478              * CALCULATE INTERACTIONS *
1479              **************************/
1480
1481             if (gmx_mm_any_lt(rsq12,rcutoff2))
1482             {
1483
1484             /* REACTION-FIELD ELECTROSTATICS */
1485             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1486
1487             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1488
1489             fscal            = felec;
1490
1491             fscal            = _mm_and_pd(fscal,cutoff_mask);
1492
1493             /* Update vectorial force */
1494             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1495             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1496             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1497
1498             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1499             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1500             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1501
1502             }
1503
1504             /**************************
1505              * CALCULATE INTERACTIONS *
1506              **************************/
1507
1508             if (gmx_mm_any_lt(rsq13,rcutoff2))
1509             {
1510
1511             /* REACTION-FIELD ELECTROSTATICS */
1512             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1513
1514             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1515
1516             fscal            = felec;
1517
1518             fscal            = _mm_and_pd(fscal,cutoff_mask);
1519
1520             /* Update vectorial force */
1521             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1522             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1523             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1524
1525             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1526             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1527             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1528
1529             }
1530
1531             /**************************
1532              * CALCULATE INTERACTIONS *
1533              **************************/
1534
1535             if (gmx_mm_any_lt(rsq21,rcutoff2))
1536             {
1537
1538             /* REACTION-FIELD ELECTROSTATICS */
1539             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1540
1541             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1542
1543             fscal            = felec;
1544
1545             fscal            = _mm_and_pd(fscal,cutoff_mask);
1546
1547             /* Update vectorial force */
1548             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1549             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1550             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1551
1552             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1553             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1554             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1555
1556             }
1557
1558             /**************************
1559              * CALCULATE INTERACTIONS *
1560              **************************/
1561
1562             if (gmx_mm_any_lt(rsq22,rcutoff2))
1563             {
1564
1565             /* REACTION-FIELD ELECTROSTATICS */
1566             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1567
1568             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1569
1570             fscal            = felec;
1571
1572             fscal            = _mm_and_pd(fscal,cutoff_mask);
1573
1574             /* Update vectorial force */
1575             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1576             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1577             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1578
1579             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1580             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1581             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1582
1583             }
1584
1585             /**************************
1586              * CALCULATE INTERACTIONS *
1587              **************************/
1588
1589             if (gmx_mm_any_lt(rsq23,rcutoff2))
1590             {
1591
1592             /* REACTION-FIELD ELECTROSTATICS */
1593             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1594
1595             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1596
1597             fscal            = felec;
1598
1599             fscal            = _mm_and_pd(fscal,cutoff_mask);
1600
1601             /* Update vectorial force */
1602             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1603             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1604             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1605
1606             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1607             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1608             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1609
1610             }
1611
1612             /**************************
1613              * CALCULATE INTERACTIONS *
1614              **************************/
1615
1616             if (gmx_mm_any_lt(rsq31,rcutoff2))
1617             {
1618
1619             /* REACTION-FIELD ELECTROSTATICS */
1620             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1621
1622             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1623
1624             fscal            = felec;
1625
1626             fscal            = _mm_and_pd(fscal,cutoff_mask);
1627
1628             /* Update vectorial force */
1629             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1630             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1631             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1632
1633             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1634             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1635             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1636
1637             }
1638
1639             /**************************
1640              * CALCULATE INTERACTIONS *
1641              **************************/
1642
1643             if (gmx_mm_any_lt(rsq32,rcutoff2))
1644             {
1645
1646             /* REACTION-FIELD ELECTROSTATICS */
1647             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1648
1649             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1650
1651             fscal            = felec;
1652
1653             fscal            = _mm_and_pd(fscal,cutoff_mask);
1654
1655             /* Update vectorial force */
1656             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1657             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1658             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1659
1660             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1661             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1662             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1663
1664             }
1665
1666             /**************************
1667              * CALCULATE INTERACTIONS *
1668              **************************/
1669
1670             if (gmx_mm_any_lt(rsq33,rcutoff2))
1671             {
1672
1673             /* REACTION-FIELD ELECTROSTATICS */
1674             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1675
1676             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1677
1678             fscal            = felec;
1679
1680             fscal            = _mm_and_pd(fscal,cutoff_mask);
1681
1682             /* Update vectorial force */
1683             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1684             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1685             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1686
1687             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1688             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1689             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1690
1691             }
1692
1693             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);
1694
1695             /* Inner loop uses 359 flops */
1696         }
1697
1698         if(jidx<j_index_end)
1699         {
1700
1701             jnrA             = jjnr[jidx];
1702             j_coord_offsetA  = DIM*jnrA;
1703
1704             /* load j atom coordinates */
1705             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1706                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1707                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1708
1709             /* Calculate displacement vector */
1710             dx00             = _mm_sub_pd(ix0,jx0);
1711             dy00             = _mm_sub_pd(iy0,jy0);
1712             dz00             = _mm_sub_pd(iz0,jz0);
1713             dx11             = _mm_sub_pd(ix1,jx1);
1714             dy11             = _mm_sub_pd(iy1,jy1);
1715             dz11             = _mm_sub_pd(iz1,jz1);
1716             dx12             = _mm_sub_pd(ix1,jx2);
1717             dy12             = _mm_sub_pd(iy1,jy2);
1718             dz12             = _mm_sub_pd(iz1,jz2);
1719             dx13             = _mm_sub_pd(ix1,jx3);
1720             dy13             = _mm_sub_pd(iy1,jy3);
1721             dz13             = _mm_sub_pd(iz1,jz3);
1722             dx21             = _mm_sub_pd(ix2,jx1);
1723             dy21             = _mm_sub_pd(iy2,jy1);
1724             dz21             = _mm_sub_pd(iz2,jz1);
1725             dx22             = _mm_sub_pd(ix2,jx2);
1726             dy22             = _mm_sub_pd(iy2,jy2);
1727             dz22             = _mm_sub_pd(iz2,jz2);
1728             dx23             = _mm_sub_pd(ix2,jx3);
1729             dy23             = _mm_sub_pd(iy2,jy3);
1730             dz23             = _mm_sub_pd(iz2,jz3);
1731             dx31             = _mm_sub_pd(ix3,jx1);
1732             dy31             = _mm_sub_pd(iy3,jy1);
1733             dz31             = _mm_sub_pd(iz3,jz1);
1734             dx32             = _mm_sub_pd(ix3,jx2);
1735             dy32             = _mm_sub_pd(iy3,jy2);
1736             dz32             = _mm_sub_pd(iz3,jz2);
1737             dx33             = _mm_sub_pd(ix3,jx3);
1738             dy33             = _mm_sub_pd(iy3,jy3);
1739             dz33             = _mm_sub_pd(iz3,jz3);
1740
1741             /* Calculate squared distance and things based on it */
1742             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1743             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1744             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1745             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1746             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1747             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1748             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1749             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1750             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1751             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1752
1753             rinv00           = avx128fma_invsqrt_d(rsq00);
1754             rinv11           = avx128fma_invsqrt_d(rsq11);
1755             rinv12           = avx128fma_invsqrt_d(rsq12);
1756             rinv13           = avx128fma_invsqrt_d(rsq13);
1757             rinv21           = avx128fma_invsqrt_d(rsq21);
1758             rinv22           = avx128fma_invsqrt_d(rsq22);
1759             rinv23           = avx128fma_invsqrt_d(rsq23);
1760             rinv31           = avx128fma_invsqrt_d(rsq31);
1761             rinv32           = avx128fma_invsqrt_d(rsq32);
1762             rinv33           = avx128fma_invsqrt_d(rsq33);
1763
1764             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1765             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1766             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1767             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1768             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1769             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1770             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1771             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1772             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1773             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1774
1775             fjx0             = _mm_setzero_pd();
1776             fjy0             = _mm_setzero_pd();
1777             fjz0             = _mm_setzero_pd();
1778             fjx1             = _mm_setzero_pd();
1779             fjy1             = _mm_setzero_pd();
1780             fjz1             = _mm_setzero_pd();
1781             fjx2             = _mm_setzero_pd();
1782             fjy2             = _mm_setzero_pd();
1783             fjz2             = _mm_setzero_pd();
1784             fjx3             = _mm_setzero_pd();
1785             fjy3             = _mm_setzero_pd();
1786             fjz3             = _mm_setzero_pd();
1787
1788             /**************************
1789              * CALCULATE INTERACTIONS *
1790              **************************/
1791
1792             if (gmx_mm_any_lt(rsq00,rcutoff2))
1793             {
1794
1795             r00              = _mm_mul_pd(rsq00,rinv00);
1796
1797             /* LENNARD-JONES DISPERSION/REPULSION */
1798
1799             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1800             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
1801             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
1802             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
1803             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
1804
1805             d                = _mm_sub_pd(r00,rswitch);
1806             d                = _mm_max_pd(d,_mm_setzero_pd());
1807             d2               = _mm_mul_pd(d,d);
1808             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
1809
1810             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
1811
1812             /* Evaluate switch function */
1813             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1814             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
1815             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1816
1817             fscal            = fvdw;
1818
1819             fscal            = _mm_and_pd(fscal,cutoff_mask);
1820
1821             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1822
1823             /* Update vectorial force */
1824             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1825             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1826             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1827
1828             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1829             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1830             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1831
1832             }
1833
1834             /**************************
1835              * CALCULATE INTERACTIONS *
1836              **************************/
1837
1838             if (gmx_mm_any_lt(rsq11,rcutoff2))
1839             {
1840
1841             /* REACTION-FIELD ELECTROSTATICS */
1842             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1843
1844             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1845
1846             fscal            = felec;
1847
1848             fscal            = _mm_and_pd(fscal,cutoff_mask);
1849
1850             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1851
1852             /* Update vectorial force */
1853             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1854             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1855             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1856
1857             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1858             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1859             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1860
1861             }
1862
1863             /**************************
1864              * CALCULATE INTERACTIONS *
1865              **************************/
1866
1867             if (gmx_mm_any_lt(rsq12,rcutoff2))
1868             {
1869
1870             /* REACTION-FIELD ELECTROSTATICS */
1871             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1872
1873             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1874
1875             fscal            = felec;
1876
1877             fscal            = _mm_and_pd(fscal,cutoff_mask);
1878
1879             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1880
1881             /* Update vectorial force */
1882             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1883             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1884             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1885
1886             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1887             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1888             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1889
1890             }
1891
1892             /**************************
1893              * CALCULATE INTERACTIONS *
1894              **************************/
1895
1896             if (gmx_mm_any_lt(rsq13,rcutoff2))
1897             {
1898
1899             /* REACTION-FIELD ELECTROSTATICS */
1900             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1901
1902             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1903
1904             fscal            = felec;
1905
1906             fscal            = _mm_and_pd(fscal,cutoff_mask);
1907
1908             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1909
1910             /* Update vectorial force */
1911             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1912             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1913             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1914
1915             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1916             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1917             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1918
1919             }
1920
1921             /**************************
1922              * CALCULATE INTERACTIONS *
1923              **************************/
1924
1925             if (gmx_mm_any_lt(rsq21,rcutoff2))
1926             {
1927
1928             /* REACTION-FIELD ELECTROSTATICS */
1929             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1930
1931             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1932
1933             fscal            = felec;
1934
1935             fscal            = _mm_and_pd(fscal,cutoff_mask);
1936
1937             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1938
1939             /* Update vectorial force */
1940             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1941             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1942             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1943
1944             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1945             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1946             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1947
1948             }
1949
1950             /**************************
1951              * CALCULATE INTERACTIONS *
1952              **************************/
1953
1954             if (gmx_mm_any_lt(rsq22,rcutoff2))
1955             {
1956
1957             /* REACTION-FIELD ELECTROSTATICS */
1958             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1959
1960             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1961
1962             fscal            = felec;
1963
1964             fscal            = _mm_and_pd(fscal,cutoff_mask);
1965
1966             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1967
1968             /* Update vectorial force */
1969             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1970             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1971             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1972
1973             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1974             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1975             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1976
1977             }
1978
1979             /**************************
1980              * CALCULATE INTERACTIONS *
1981              **************************/
1982
1983             if (gmx_mm_any_lt(rsq23,rcutoff2))
1984             {
1985
1986             /* REACTION-FIELD ELECTROSTATICS */
1987             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1988
1989             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1990
1991             fscal            = felec;
1992
1993             fscal            = _mm_and_pd(fscal,cutoff_mask);
1994
1995             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1996
1997             /* Update vectorial force */
1998             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1999             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
2000             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
2001
2002             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
2003             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
2004             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
2005
2006             }
2007
2008             /**************************
2009              * CALCULATE INTERACTIONS *
2010              **************************/
2011
2012             if (gmx_mm_any_lt(rsq31,rcutoff2))
2013             {
2014
2015             /* REACTION-FIELD ELECTROSTATICS */
2016             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
2017
2018             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
2019
2020             fscal            = felec;
2021
2022             fscal            = _mm_and_pd(fscal,cutoff_mask);
2023
2024             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2025
2026             /* Update vectorial force */
2027             fix3             = _mm_macc_pd(dx31,fscal,fix3);
2028             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
2029             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
2030
2031             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
2032             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
2033             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
2034
2035             }
2036
2037             /**************************
2038              * CALCULATE INTERACTIONS *
2039              **************************/
2040
2041             if (gmx_mm_any_lt(rsq32,rcutoff2))
2042             {
2043
2044             /* REACTION-FIELD ELECTROSTATICS */
2045             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
2046
2047             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
2048
2049             fscal            = felec;
2050
2051             fscal            = _mm_and_pd(fscal,cutoff_mask);
2052
2053             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2054
2055             /* Update vectorial force */
2056             fix3             = _mm_macc_pd(dx32,fscal,fix3);
2057             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
2058             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
2059
2060             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
2061             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
2062             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
2063
2064             }
2065
2066             /**************************
2067              * CALCULATE INTERACTIONS *
2068              **************************/
2069
2070             if (gmx_mm_any_lt(rsq33,rcutoff2))
2071             {
2072
2073             /* REACTION-FIELD ELECTROSTATICS */
2074             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
2075
2076             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
2077
2078             fscal            = felec;
2079
2080             fscal            = _mm_and_pd(fscal,cutoff_mask);
2081
2082             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2083
2084             /* Update vectorial force */
2085             fix3             = _mm_macc_pd(dx33,fscal,fix3);
2086             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
2087             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
2088
2089             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
2090             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
2091             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
2092
2093             }
2094
2095             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2096
2097             /* Inner loop uses 359 flops */
2098         }
2099
2100         /* End of innermost loop */
2101
2102         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2103                                               f+i_coord_offset,fshift+i_shift_offset);
2104
2105         /* Increment number of inner iterations */
2106         inneriter                  += j_index_end - j_index_start;
2107
2108         /* Outer loop uses 24 flops */
2109     }
2110
2111     /* Increment number of outer iterations */
2112     outeriter        += nri;
2113
2114     /* Update outer/inner flops */
2115
2116     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*359);
2117 }