src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_double/nb_kernel_ElecRFCut_VdwCSTab_GeomW4W4_sse2_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017, 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 sse2_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_sse2_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW4W4_VF_sse2_double
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            CubicSplineTable
  53  * Geometry:                   Water4-Water4
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRFCut_VdwCSTab_GeomW4W4_VF_sse2_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     __m128i          vfitab;
 114     __m128i          ifour       = _mm_set1_epi32(4);
 115     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 116     real             *vftab;
 117     __m128d          dummy_mask,cutoff_mask;
 118     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 119     __m128d          one     = _mm_set1_pd(1.0);
 120     __m128d          two     = _mm_set1_pd(2.0);
 121     x                = xx[0];
 122     f                = ff[0];
 123
 124     nri              = nlist->nri;
 125     iinr             = nlist->iinr;
 126     jindex           = nlist->jindex;
 127     jjnr             = nlist->jjnr;
 128     shiftidx         = nlist->shift;
 129     gid              = nlist->gid;
 130     shiftvec         = fr->shift_vec[0];
 131     fshift           = fr->fshift[0];
 132     facel            = _mm_set1_pd(fr->ic->epsfac);
 133     charge           = mdatoms->chargeA;
 134     krf              = _mm_set1_pd(fr->ic->k_rf);
 135     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 136     crf              = _mm_set1_pd(fr->ic->c_rf);
 137     nvdwtype         = fr->ntype;
 138     vdwparam         = fr->nbfp;
 139     vdwtype          = mdatoms->typeA;
 140
 141     vftab            = kernel_data->table_vdw->data;
 142     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
 143
 144     /* Setup water-specific parameters */
 145     inr              = nlist->iinr[0];
 146     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 147     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 148     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 149     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 150
 151     jq1              = _mm_set1_pd(charge[inr+1]);
 152     jq2              = _mm_set1_pd(charge[inr+2]);
 153     jq3              = _mm_set1_pd(charge[inr+3]);
 154     vdwjidx0A        = 2*vdwtype[inr+0];
 155     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 156     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 157     qq11             = _mm_mul_pd(iq1,jq1);
 158     qq12             = _mm_mul_pd(iq1,jq2);
 159     qq13             = _mm_mul_pd(iq1,jq3);
 160     qq21             = _mm_mul_pd(iq2,jq1);
 161     qq22             = _mm_mul_pd(iq2,jq2);
 162     qq23             = _mm_mul_pd(iq2,jq3);
 163     qq31             = _mm_mul_pd(iq3,jq1);
 164     qq32             = _mm_mul_pd(iq3,jq2);
 165     qq33             = _mm_mul_pd(iq3,jq3);
 166
 167     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 168     rcutoff_scalar   = fr->ic->rcoulomb;
 169     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 170     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 171
 172     /* Avoid stupid compiler warnings */
 173     jnrA = jnrB = 0;
 174     j_coord_offsetA = 0;
 175     j_coord_offsetB = 0;
 176
 177     outeriter        = 0;
 178     inneriter        = 0;
 179
 180     /* Start outer loop over neighborlists */
 181     for(iidx=0; iidx<nri; iidx++)
 182     {
 183         /* Load shift vector for this list */
 184         i_shift_offset   = DIM*shiftidx[iidx];
 185
 186         /* Load limits for loop over neighbors */
 187         j_index_start    = jindex[iidx];
 188         j_index_end      = jindex[iidx+1];
 189
 190         /* Get outer coordinate index */
 191         inr              = iinr[iidx];
 192         i_coord_offset   = DIM*inr;
 193
 194         /* Load i particle coords and add shift vector */
 195         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 196                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 197
 198         fix0             = _mm_setzero_pd();
 199         fiy0             = _mm_setzero_pd();
 200         fiz0             = _mm_setzero_pd();
 201         fix1             = _mm_setzero_pd();
 202         fiy1             = _mm_setzero_pd();
 203         fiz1             = _mm_setzero_pd();
 204         fix2             = _mm_setzero_pd();
 205         fiy2             = _mm_setzero_pd();
 206         fiz2             = _mm_setzero_pd();
 207         fix3             = _mm_setzero_pd();
 208         fiy3             = _mm_setzero_pd();
 209         fiz3             = _mm_setzero_pd();
 210
 211         /* Reset potential sums */
 212         velecsum         = _mm_setzero_pd();
 213         vvdwsum          = _mm_setzero_pd();
 214
 215         /* Start inner kernel loop */
 216         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 217         {
 218
 219             /* Get j neighbor index, and coordinate index */
 220             jnrA             = jjnr[jidx];
 221             jnrB             = jjnr[jidx+1];
 222             j_coord_offsetA  = DIM*jnrA;
 223             j_coord_offsetB  = DIM*jnrB;
 224
 225             /* load j atom coordinates */
 226             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 227                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 228                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 229
 230             /* Calculate displacement vector */
 231             dx00             = _mm_sub_pd(ix0,jx0);
 232             dy00             = _mm_sub_pd(iy0,jy0);
 233             dz00             = _mm_sub_pd(iz0,jz0);
 234             dx11             = _mm_sub_pd(ix1,jx1);
 235             dy11             = _mm_sub_pd(iy1,jy1);
 236             dz11             = _mm_sub_pd(iz1,jz1);
 237             dx12             = _mm_sub_pd(ix1,jx2);
 238             dy12             = _mm_sub_pd(iy1,jy2);
 239             dz12             = _mm_sub_pd(iz1,jz2);
 240             dx13             = _mm_sub_pd(ix1,jx3);
 241             dy13             = _mm_sub_pd(iy1,jy3);
 242             dz13             = _mm_sub_pd(iz1,jz3);
 243             dx21             = _mm_sub_pd(ix2,jx1);
 244             dy21             = _mm_sub_pd(iy2,jy1);
 245             dz21             = _mm_sub_pd(iz2,jz1);
 246             dx22             = _mm_sub_pd(ix2,jx2);
 247             dy22             = _mm_sub_pd(iy2,jy2);
 248             dz22             = _mm_sub_pd(iz2,jz2);
 249             dx23             = _mm_sub_pd(ix2,jx3);
 250             dy23             = _mm_sub_pd(iy2,jy3);
 251             dz23             = _mm_sub_pd(iz2,jz3);
 252             dx31             = _mm_sub_pd(ix3,jx1);
 253             dy31             = _mm_sub_pd(iy3,jy1);
 254             dz31             = _mm_sub_pd(iz3,jz1);
 255             dx32             = _mm_sub_pd(ix3,jx2);
 256             dy32             = _mm_sub_pd(iy3,jy2);
 257             dz32             = _mm_sub_pd(iz3,jz2);
 258             dx33             = _mm_sub_pd(ix3,jx3);
 259             dy33             = _mm_sub_pd(iy3,jy3);
 260             dz33             = _mm_sub_pd(iz3,jz3);
 261
 262             /* Calculate squared distance and things based on it */
 263             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 264             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 265             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 266             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 267             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 268             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 269             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 270             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 271             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 272             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 273
 274             rinv00           = sse2_invsqrt_d(rsq00);
 275             rinv11           = sse2_invsqrt_d(rsq11);
 276             rinv12           = sse2_invsqrt_d(rsq12);
 277             rinv13           = sse2_invsqrt_d(rsq13);
 278             rinv21           = sse2_invsqrt_d(rsq21);
 279             rinv22           = sse2_invsqrt_d(rsq22);
 280             rinv23           = sse2_invsqrt_d(rsq23);
 281             rinv31           = sse2_invsqrt_d(rsq31);
 282             rinv32           = sse2_invsqrt_d(rsq32);
 283             rinv33           = sse2_invsqrt_d(rsq33);
 284
 285             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 286             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 287             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 288             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 289             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 290             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 291             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 292             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 293             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 294
 295             fjx0             = _mm_setzero_pd();
 296             fjy0             = _mm_setzero_pd();
 297             fjz0             = _mm_setzero_pd();
 298             fjx1             = _mm_setzero_pd();
 299             fjy1             = _mm_setzero_pd();
 300             fjz1             = _mm_setzero_pd();
 301             fjx2             = _mm_setzero_pd();
 302             fjy2             = _mm_setzero_pd();
 303             fjz2             = _mm_setzero_pd();
 304             fjx3             = _mm_setzero_pd();
 305             fjy3             = _mm_setzero_pd();
 306             fjz3             = _mm_setzero_pd();
 307
 308             /**************************
 309              * CALCULATE INTERACTIONS *
 310              **************************/
 311
 312             r00              = _mm_mul_pd(rsq00,rinv00);
 313
 314             /* Calculate table index by multiplying r with table scale and truncate to integer */
 315             rt               = _mm_mul_pd(r00,vftabscale);
 316             vfitab           = _mm_cvttpd_epi32(rt);
 317             vfeps            = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
 318             vfitab           = _mm_slli_epi32(vfitab,3);
 319
 320             /* CUBIC SPLINE TABLE DISPERSION */
 321             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 322             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
 323             GMX_MM_TRANSPOSE2_PD(Y,F);
 324             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 325             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
 326             GMX_MM_TRANSPOSE2_PD(G,H);
 327             Heps             = _mm_mul_pd(vfeps,H);
 328             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 329             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 330             vvdw6            = _mm_mul_pd(c6_00,VV);
 331             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 332             fvdw6            = _mm_mul_pd(c6_00,FF);
 333
 334             /* CUBIC SPLINE TABLE REPULSION */
 335             vfitab           = _mm_add_epi32(vfitab,ifour);
 336             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 337             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
 338             GMX_MM_TRANSPOSE2_PD(Y,F);
 339             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 340             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
 341             GMX_MM_TRANSPOSE2_PD(G,H);
 342             Heps             = _mm_mul_pd(vfeps,H);
 343             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 344             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 345             vvdw12           = _mm_mul_pd(c12_00,VV);
 346             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 347             fvdw12           = _mm_mul_pd(c12_00,FF);
 348             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 349             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 350
 351             /* Update potential sum for this i atom from the interaction with this j atom. */
 352             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 353
 354             fscal            = fvdw;
 355
 356             /* Calculate temporary vectorial force */
 357             tx               = _mm_mul_pd(fscal,dx00);
 358             ty               = _mm_mul_pd(fscal,dy00);
 359             tz               = _mm_mul_pd(fscal,dz00);
 360
 361             /* Update vectorial force */
 362             fix0             = _mm_add_pd(fix0,tx);
 363             fiy0             = _mm_add_pd(fiy0,ty);
 364             fiz0             = _mm_add_pd(fiz0,tz);
 365
 366             fjx0             = _mm_add_pd(fjx0,tx);
 367             fjy0             = _mm_add_pd(fjy0,ty);
 368             fjz0             = _mm_add_pd(fjz0,tz);
 369
 370             /**************************
 371              * CALCULATE INTERACTIONS *
 372              **************************/
 373
 374             if (gmx_mm_any_lt(rsq11,rcutoff2))
 375             {
 376
 377             /* REACTION-FIELD ELECTROSTATICS */
 378             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 379             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 380
 381             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 382
 383             /* Update potential sum for this i atom from the interaction with this j atom. */
 384             velec            = _mm_and_pd(velec,cutoff_mask);
 385             velecsum         = _mm_add_pd(velecsum,velec);
 386
 387             fscal            = felec;
 388
 389             fscal            = _mm_and_pd(fscal,cutoff_mask);
 390
 391             /* Calculate temporary vectorial force */
 392             tx               = _mm_mul_pd(fscal,dx11);
 393             ty               = _mm_mul_pd(fscal,dy11);
 394             tz               = _mm_mul_pd(fscal,dz11);
 395
 396             /* Update vectorial force */
 397             fix1             = _mm_add_pd(fix1,tx);
 398             fiy1             = _mm_add_pd(fiy1,ty);
 399             fiz1             = _mm_add_pd(fiz1,tz);
 400
 401             fjx1             = _mm_add_pd(fjx1,tx);
 402             fjy1             = _mm_add_pd(fjy1,ty);
 403             fjz1             = _mm_add_pd(fjz1,tz);
 404
 405             }
 406
 407             /**************************
 408              * CALCULATE INTERACTIONS *
 409              **************************/
 410
 411             if (gmx_mm_any_lt(rsq12,rcutoff2))
 412             {
 413
 414             /* REACTION-FIELD ELECTROSTATICS */
 415             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 416             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 417
 418             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 419
 420             /* Update potential sum for this i atom from the interaction with this j atom. */
 421             velec            = _mm_and_pd(velec,cutoff_mask);
 422             velecsum         = _mm_add_pd(velecsum,velec);
 423
 424             fscal            = felec;
 425
 426             fscal            = _mm_and_pd(fscal,cutoff_mask);
 427
 428             /* Calculate temporary vectorial force */
 429             tx               = _mm_mul_pd(fscal,dx12);
 430             ty               = _mm_mul_pd(fscal,dy12);
 431             tz               = _mm_mul_pd(fscal,dz12);
 432
 433             /* Update vectorial force */
 434             fix1             = _mm_add_pd(fix1,tx);
 435             fiy1             = _mm_add_pd(fiy1,ty);
 436             fiz1             = _mm_add_pd(fiz1,tz);
 437
 438             fjx2             = _mm_add_pd(fjx2,tx);
 439             fjy2             = _mm_add_pd(fjy2,ty);
 440             fjz2             = _mm_add_pd(fjz2,tz);
 441
 442             }
 443
 444             /**************************
 445              * CALCULATE INTERACTIONS *
 446              **************************/
 447
 448             if (gmx_mm_any_lt(rsq13,rcutoff2))
 449             {
 450
 451             /* REACTION-FIELD ELECTROSTATICS */
 452             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
 453             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
 454
 455             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 456
 457             /* Update potential sum for this i atom from the interaction with this j atom. */
 458             velec            = _mm_and_pd(velec,cutoff_mask);
 459             velecsum         = _mm_add_pd(velecsum,velec);
 460
 461             fscal            = felec;
 462
 463             fscal            = _mm_and_pd(fscal,cutoff_mask);
 464
 465             /* Calculate temporary vectorial force */
 466             tx               = _mm_mul_pd(fscal,dx13);
 467             ty               = _mm_mul_pd(fscal,dy13);
 468             tz               = _mm_mul_pd(fscal,dz13);
 469
 470             /* Update vectorial force */
 471             fix1             = _mm_add_pd(fix1,tx);
 472             fiy1             = _mm_add_pd(fiy1,ty);
 473             fiz1             = _mm_add_pd(fiz1,tz);
 474
 475             fjx3             = _mm_add_pd(fjx3,tx);
 476             fjy3             = _mm_add_pd(fjy3,ty);
 477             fjz3             = _mm_add_pd(fjz3,tz);
 478
 479             }
 480
 481             /**************************
 482              * CALCULATE INTERACTIONS *
 483              **************************/
 484
 485             if (gmx_mm_any_lt(rsq21,rcutoff2))
 486             {
 487
 488             /* REACTION-FIELD ELECTROSTATICS */
 489             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 490             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 491
 492             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 493
 494             /* Update potential sum for this i atom from the interaction with this j atom. */
 495             velec            = _mm_and_pd(velec,cutoff_mask);
 496             velecsum         = _mm_add_pd(velecsum,velec);
 497
 498             fscal            = felec;
 499
 500             fscal            = _mm_and_pd(fscal,cutoff_mask);
 501
 502             /* Calculate temporary vectorial force */
 503             tx               = _mm_mul_pd(fscal,dx21);
 504             ty               = _mm_mul_pd(fscal,dy21);
 505             tz               = _mm_mul_pd(fscal,dz21);
 506
 507             /* Update vectorial force */
 508             fix2             = _mm_add_pd(fix2,tx);
 509             fiy2             = _mm_add_pd(fiy2,ty);
 510             fiz2             = _mm_add_pd(fiz2,tz);
 511
 512             fjx1             = _mm_add_pd(fjx1,tx);
 513             fjy1             = _mm_add_pd(fjy1,ty);
 514             fjz1             = _mm_add_pd(fjz1,tz);
 515
 516             }
 517
 518             /**************************
 519              * CALCULATE INTERACTIONS *
 520              **************************/
 521
 522             if (gmx_mm_any_lt(rsq22,rcutoff2))
 523             {
 524
 525             /* REACTION-FIELD ELECTROSTATICS */
 526             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 527             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 528
 529             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 530
 531             /* Update potential sum for this i atom from the interaction with this j atom. */
 532             velec            = _mm_and_pd(velec,cutoff_mask);
 533             velecsum         = _mm_add_pd(velecsum,velec);
 534
 535             fscal            = felec;
 536
 537             fscal            = _mm_and_pd(fscal,cutoff_mask);
 538
 539             /* Calculate temporary vectorial force */
 540             tx               = _mm_mul_pd(fscal,dx22);
 541             ty               = _mm_mul_pd(fscal,dy22);
 542             tz               = _mm_mul_pd(fscal,dz22);
 543
 544             /* Update vectorial force */
 545             fix2             = _mm_add_pd(fix2,tx);
 546             fiy2             = _mm_add_pd(fiy2,ty);
 547             fiz2             = _mm_add_pd(fiz2,tz);
 548
 549             fjx2             = _mm_add_pd(fjx2,tx);
 550             fjy2             = _mm_add_pd(fjy2,ty);
 551             fjz2             = _mm_add_pd(fjz2,tz);
 552
 553             }
 554
 555             /**************************
 556              * CALCULATE INTERACTIONS *
 557              **************************/
 558
 559             if (gmx_mm_any_lt(rsq23,rcutoff2))
 560             {
 561
 562             /* REACTION-FIELD ELECTROSTATICS */
 563             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
 564             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
 565
 566             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 567
 568             /* Update potential sum for this i atom from the interaction with this j atom. */
 569             velec            = _mm_and_pd(velec,cutoff_mask);
 570             velecsum         = _mm_add_pd(velecsum,velec);
 571
 572             fscal            = felec;
 573
 574             fscal            = _mm_and_pd(fscal,cutoff_mask);
 575
 576             /* Calculate temporary vectorial force */
 577             tx               = _mm_mul_pd(fscal,dx23);
 578             ty               = _mm_mul_pd(fscal,dy23);
 579             tz               = _mm_mul_pd(fscal,dz23);
 580
 581             /* Update vectorial force */
 582             fix2             = _mm_add_pd(fix2,tx);
 583             fiy2             = _mm_add_pd(fiy2,ty);
 584             fiz2             = _mm_add_pd(fiz2,tz);
 585
 586             fjx3             = _mm_add_pd(fjx3,tx);
 587             fjy3             = _mm_add_pd(fjy3,ty);
 588             fjz3             = _mm_add_pd(fjz3,tz);
 589
 590             }
 591
 592             /**************************
 593              * CALCULATE INTERACTIONS *
 594              **************************/
 595
 596             if (gmx_mm_any_lt(rsq31,rcutoff2))
 597             {
 598
 599             /* REACTION-FIELD ELECTROSTATICS */
 600             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
 601             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
 602
 603             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
 604
 605             /* Update potential sum for this i atom from the interaction with this j atom. */
 606             velec            = _mm_and_pd(velec,cutoff_mask);
 607             velecsum         = _mm_add_pd(velecsum,velec);
 608
 609             fscal            = felec;
 610
 611             fscal            = _mm_and_pd(fscal,cutoff_mask);
 612
 613             /* Calculate temporary vectorial force */
 614             tx               = _mm_mul_pd(fscal,dx31);
 615             ty               = _mm_mul_pd(fscal,dy31);
 616             tz               = _mm_mul_pd(fscal,dz31);
 617
 618             /* Update vectorial force */
 619             fix3             = _mm_add_pd(fix3,tx);
 620             fiy3             = _mm_add_pd(fiy3,ty);
 621             fiz3             = _mm_add_pd(fiz3,tz);
 622
 623             fjx1             = _mm_add_pd(fjx1,tx);
 624             fjy1             = _mm_add_pd(fjy1,ty);
 625             fjz1             = _mm_add_pd(fjz1,tz);
 626
 627             }
 628
 629             /**************************
 630              * CALCULATE INTERACTIONS *
 631              **************************/
 632
 633             if (gmx_mm_any_lt(rsq32,rcutoff2))
 634             {
 635
 636             /* REACTION-FIELD ELECTROSTATICS */
 637             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
 638             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
 639
 640             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
 641
 642             /* Update potential sum for this i atom from the interaction with this j atom. */
 643             velec            = _mm_and_pd(velec,cutoff_mask);
 644             velecsum         = _mm_add_pd(velecsum,velec);
 645
 646             fscal            = felec;
 647
 648             fscal            = _mm_and_pd(fscal,cutoff_mask);
 649
 650             /* Calculate temporary vectorial force */
 651             tx               = _mm_mul_pd(fscal,dx32);
 652             ty               = _mm_mul_pd(fscal,dy32);
 653             tz               = _mm_mul_pd(fscal,dz32);
 654
 655             /* Update vectorial force */
 656             fix3             = _mm_add_pd(fix3,tx);
 657             fiy3             = _mm_add_pd(fiy3,ty);
 658             fiz3             = _mm_add_pd(fiz3,tz);
 659
 660             fjx2             = _mm_add_pd(fjx2,tx);
 661             fjy2             = _mm_add_pd(fjy2,ty);
 662             fjz2             = _mm_add_pd(fjz2,tz);
 663
 664             }
 665
 666             /**************************
 667              * CALCULATE INTERACTIONS *
 668              **************************/
 669
 670             if (gmx_mm_any_lt(rsq33,rcutoff2))
 671             {
 672
 673             /* REACTION-FIELD ELECTROSTATICS */
 674             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
 675             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
 676
 677             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
 678
 679             /* Update potential sum for this i atom from the interaction with this j atom. */
 680             velec            = _mm_and_pd(velec,cutoff_mask);
 681             velecsum         = _mm_add_pd(velecsum,velec);
 682
 683             fscal            = felec;
 684
 685             fscal            = _mm_and_pd(fscal,cutoff_mask);
 686
 687             /* Calculate temporary vectorial force */
 688             tx               = _mm_mul_pd(fscal,dx33);
 689             ty               = _mm_mul_pd(fscal,dy33);
 690             tz               = _mm_mul_pd(fscal,dz33);
 691
 692             /* Update vectorial force */
 693             fix3             = _mm_add_pd(fix3,tx);
 694             fiy3             = _mm_add_pd(fiy3,ty);
 695             fiz3             = _mm_add_pd(fiz3,tz);
 696
 697             fjx3             = _mm_add_pd(fjx3,tx);
 698             fjy3             = _mm_add_pd(fjy3,ty);
 699             fjz3             = _mm_add_pd(fjz3,tz);
 700
 701             }
 702
 703             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);
 704
 705             /* Inner loop uses 383 flops */
 706         }
 707
 708         if(jidx<j_index_end)
 709         {
 710
 711             jnrA             = jjnr[jidx];
 712             j_coord_offsetA  = DIM*jnrA;
 713
 714             /* load j atom coordinates */
 715             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 716                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 717                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 718
 719             /* Calculate displacement vector */
 720             dx00             = _mm_sub_pd(ix0,jx0);
 721             dy00             = _mm_sub_pd(iy0,jy0);
 722             dz00             = _mm_sub_pd(iz0,jz0);
 723             dx11             = _mm_sub_pd(ix1,jx1);
 724             dy11             = _mm_sub_pd(iy1,jy1);
 725             dz11             = _mm_sub_pd(iz1,jz1);
 726             dx12             = _mm_sub_pd(ix1,jx2);
 727             dy12             = _mm_sub_pd(iy1,jy2);
 728             dz12             = _mm_sub_pd(iz1,jz2);
 729             dx13             = _mm_sub_pd(ix1,jx3);
 730             dy13             = _mm_sub_pd(iy1,jy3);
 731             dz13             = _mm_sub_pd(iz1,jz3);
 732             dx21             = _mm_sub_pd(ix2,jx1);
 733             dy21             = _mm_sub_pd(iy2,jy1);
 734             dz21             = _mm_sub_pd(iz2,jz1);
 735             dx22             = _mm_sub_pd(ix2,jx2);
 736             dy22             = _mm_sub_pd(iy2,jy2);
 737             dz22             = _mm_sub_pd(iz2,jz2);
 738             dx23             = _mm_sub_pd(ix2,jx3);
 739             dy23             = _mm_sub_pd(iy2,jy3);
 740             dz23             = _mm_sub_pd(iz2,jz3);
 741             dx31             = _mm_sub_pd(ix3,jx1);
 742             dy31             = _mm_sub_pd(iy3,jy1);
 743             dz31             = _mm_sub_pd(iz3,jz1);
 744             dx32             = _mm_sub_pd(ix3,jx2);
 745             dy32             = _mm_sub_pd(iy3,jy2);
 746             dz32             = _mm_sub_pd(iz3,jz2);
 747             dx33             = _mm_sub_pd(ix3,jx3);
 748             dy33             = _mm_sub_pd(iy3,jy3);
 749             dz33             = _mm_sub_pd(iz3,jz3);
 750
 751             /* Calculate squared distance and things based on it */
 752             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 753             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 754             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 755             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 756             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 757             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 758             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 759             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 760             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 761             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 762
 763             rinv00           = sse2_invsqrt_d(rsq00);
 764             rinv11           = sse2_invsqrt_d(rsq11);
 765             rinv12           = sse2_invsqrt_d(rsq12);
 766             rinv13           = sse2_invsqrt_d(rsq13);
 767             rinv21           = sse2_invsqrt_d(rsq21);
 768             rinv22           = sse2_invsqrt_d(rsq22);
 769             rinv23           = sse2_invsqrt_d(rsq23);
 770             rinv31           = sse2_invsqrt_d(rsq31);
 771             rinv32           = sse2_invsqrt_d(rsq32);
 772             rinv33           = sse2_invsqrt_d(rsq33);
 773
 774             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 775             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 776             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 777             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 778             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 779             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 780             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 781             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 782             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 783
 784             fjx0             = _mm_setzero_pd();
 785             fjy0             = _mm_setzero_pd();
 786             fjz0             = _mm_setzero_pd();
 787             fjx1             = _mm_setzero_pd();
 788             fjy1             = _mm_setzero_pd();
 789             fjz1             = _mm_setzero_pd();
 790             fjx2             = _mm_setzero_pd();
 791             fjy2             = _mm_setzero_pd();
 792             fjz2             = _mm_setzero_pd();
 793             fjx3             = _mm_setzero_pd();
 794             fjy3             = _mm_setzero_pd();
 795             fjz3             = _mm_setzero_pd();
 796
 797             /**************************
 798              * CALCULATE INTERACTIONS *
 799              **************************/
 800
 801             r00              = _mm_mul_pd(rsq00,rinv00);
 802
 803             /* Calculate table index by multiplying r with table scale and truncate to integer */
 804             rt               = _mm_mul_pd(r00,vftabscale);
 805             vfitab           = _mm_cvttpd_epi32(rt);
 806             vfeps            = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
 807             vfitab           = _mm_slli_epi32(vfitab,3);
 808
 809             /* CUBIC SPLINE TABLE DISPERSION */
 810             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 811             F                = _mm_setzero_pd();
 812             GMX_MM_TRANSPOSE2_PD(Y,F);
 813             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 814             H                = _mm_setzero_pd();
 815             GMX_MM_TRANSPOSE2_PD(G,H);
 816             Heps             = _mm_mul_pd(vfeps,H);
 817             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 818             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 819             vvdw6            = _mm_mul_pd(c6_00,VV);
 820             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 821             fvdw6            = _mm_mul_pd(c6_00,FF);
 822
 823             /* CUBIC SPLINE TABLE REPULSION */
 824             vfitab           = _mm_add_epi32(vfitab,ifour);
 825             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 826             F                = _mm_setzero_pd();
 827             GMX_MM_TRANSPOSE2_PD(Y,F);
 828             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 829             H                = _mm_setzero_pd();
 830             GMX_MM_TRANSPOSE2_PD(G,H);
 831             Heps             = _mm_mul_pd(vfeps,H);
 832             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 833             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 834             vvdw12           = _mm_mul_pd(c12_00,VV);
 835             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 836             fvdw12           = _mm_mul_pd(c12_00,FF);
 837             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 838             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 839
 840             /* Update potential sum for this i atom from the interaction with this j atom. */
 841             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 842             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 843
 844             fscal            = fvdw;
 845
 846             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 847
 848             /* Calculate temporary vectorial force */
 849             tx               = _mm_mul_pd(fscal,dx00);
 850             ty               = _mm_mul_pd(fscal,dy00);
 851             tz               = _mm_mul_pd(fscal,dz00);
 852
 853             /* Update vectorial force */
 854             fix0             = _mm_add_pd(fix0,tx);
 855             fiy0             = _mm_add_pd(fiy0,ty);
 856             fiz0             = _mm_add_pd(fiz0,tz);
 857
 858             fjx0             = _mm_add_pd(fjx0,tx);
 859             fjy0             = _mm_add_pd(fjy0,ty);
 860             fjz0             = _mm_add_pd(fjz0,tz);
 861
 862             /**************************
 863              * CALCULATE INTERACTIONS *
 864              **************************/
 865
 866             if (gmx_mm_any_lt(rsq11,rcutoff2))
 867             {
 868
 869             /* REACTION-FIELD ELECTROSTATICS */
 870             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 871             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 872
 873             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 874
 875             /* Update potential sum for this i atom from the interaction with this j atom. */
 876             velec            = _mm_and_pd(velec,cutoff_mask);
 877             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 878             velecsum         = _mm_add_pd(velecsum,velec);
 879
 880             fscal            = felec;
 881
 882             fscal            = _mm_and_pd(fscal,cutoff_mask);
 883
 884             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 885
 886             /* Calculate temporary vectorial force */
 887             tx               = _mm_mul_pd(fscal,dx11);
 888             ty               = _mm_mul_pd(fscal,dy11);
 889             tz               = _mm_mul_pd(fscal,dz11);
 890
 891             /* Update vectorial force */
 892             fix1             = _mm_add_pd(fix1,tx);
 893             fiy1             = _mm_add_pd(fiy1,ty);
 894             fiz1             = _mm_add_pd(fiz1,tz);
 895
 896             fjx1             = _mm_add_pd(fjx1,tx);
 897             fjy1             = _mm_add_pd(fjy1,ty);
 898             fjz1             = _mm_add_pd(fjz1,tz);
 899
 900             }
 901
 902             /**************************
 903              * CALCULATE INTERACTIONS *
 904              **************************/
 905
 906             if (gmx_mm_any_lt(rsq12,rcutoff2))
 907             {
 908
 909             /* REACTION-FIELD ELECTROSTATICS */
 910             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 911             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 912
 913             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 914
 915             /* Update potential sum for this i atom from the interaction with this j atom. */
 916             velec            = _mm_and_pd(velec,cutoff_mask);
 917             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 918             velecsum         = _mm_add_pd(velecsum,velec);
 919
 920             fscal            = felec;
 921
 922             fscal            = _mm_and_pd(fscal,cutoff_mask);
 923
 924             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 925
 926             /* Calculate temporary vectorial force */
 927             tx               = _mm_mul_pd(fscal,dx12);
 928             ty               = _mm_mul_pd(fscal,dy12);
 929             tz               = _mm_mul_pd(fscal,dz12);
 930
 931             /* Update vectorial force */
 932             fix1             = _mm_add_pd(fix1,tx);
 933             fiy1             = _mm_add_pd(fiy1,ty);
 934             fiz1             = _mm_add_pd(fiz1,tz);
 935
 936             fjx2             = _mm_add_pd(fjx2,tx);
 937             fjy2             = _mm_add_pd(fjy2,ty);
 938             fjz2             = _mm_add_pd(fjz2,tz);
 939
 940             }
 941
 942             /**************************
 943              * CALCULATE INTERACTIONS *
 944              **************************/
 945
 946             if (gmx_mm_any_lt(rsq13,rcutoff2))
 947             {
 948
 949             /* REACTION-FIELD ELECTROSTATICS */
 950             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
 951             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
 952
 953             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 954
 955             /* Update potential sum for this i atom from the interaction with this j atom. */
 956             velec            = _mm_and_pd(velec,cutoff_mask);
 957             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 958             velecsum         = _mm_add_pd(velecsum,velec);
 959
 960             fscal            = felec;
 961
 962             fscal            = _mm_and_pd(fscal,cutoff_mask);
 963
 964             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 965
 966             /* Calculate temporary vectorial force */
 967             tx               = _mm_mul_pd(fscal,dx13);
 968             ty               = _mm_mul_pd(fscal,dy13);
 969             tz               = _mm_mul_pd(fscal,dz13);
 970
 971             /* Update vectorial force */
 972             fix1             = _mm_add_pd(fix1,tx);
 973             fiy1             = _mm_add_pd(fiy1,ty);
 974             fiz1             = _mm_add_pd(fiz1,tz);
 975
 976             fjx3             = _mm_add_pd(fjx3,tx);
 977             fjy3             = _mm_add_pd(fjy3,ty);
 978             fjz3             = _mm_add_pd(fjz3,tz);
 979
 980             }
 981
 982             /**************************
 983              * CALCULATE INTERACTIONS *
 984              **************************/
 985
 986             if (gmx_mm_any_lt(rsq21,rcutoff2))
 987             {
 988
 989             /* REACTION-FIELD ELECTROSTATICS */
 990             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 991             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 992
 993             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 994
 995             /* Update potential sum for this i atom from the interaction with this j atom. */
 996             velec            = _mm_and_pd(velec,cutoff_mask);
 997             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 998             velecsum         = _mm_add_pd(velecsum,velec);
 999
1000             fscal            = felec;
1001
1002             fscal            = _mm_and_pd(fscal,cutoff_mask);
1003
1004             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1005
1006             /* Calculate temporary vectorial force */
1007             tx               = _mm_mul_pd(fscal,dx21);
1008             ty               = _mm_mul_pd(fscal,dy21);
1009             tz               = _mm_mul_pd(fscal,dz21);
1010
1011             /* Update vectorial force */
1012             fix2             = _mm_add_pd(fix2,tx);
1013             fiy2             = _mm_add_pd(fiy2,ty);
1014             fiz2             = _mm_add_pd(fiz2,tz);
1015
1016             fjx1             = _mm_add_pd(fjx1,tx);
1017             fjy1             = _mm_add_pd(fjy1,ty);
1018             fjz1             = _mm_add_pd(fjz1,tz);
1019
1020             }
1021
1022             /**************************
1023              * CALCULATE INTERACTIONS *
1024              **************************/
1025
1026             if (gmx_mm_any_lt(rsq22,rcutoff2))
1027             {
1028
1029             /* REACTION-FIELD ELECTROSTATICS */
1030             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
1031             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1032
1033             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1034
1035             /* Update potential sum for this i atom from the interaction with this j atom. */
1036             velec            = _mm_and_pd(velec,cutoff_mask);
1037             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1038             velecsum         = _mm_add_pd(velecsum,velec);
1039
1040             fscal            = felec;
1041
1042             fscal            = _mm_and_pd(fscal,cutoff_mask);
1043
1044             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1045
1046             /* Calculate temporary vectorial force */
1047             tx               = _mm_mul_pd(fscal,dx22);
1048             ty               = _mm_mul_pd(fscal,dy22);
1049             tz               = _mm_mul_pd(fscal,dz22);
1050
1051             /* Update vectorial force */
1052             fix2             = _mm_add_pd(fix2,tx);
1053             fiy2             = _mm_add_pd(fiy2,ty);
1054             fiz2             = _mm_add_pd(fiz2,tz);
1055
1056             fjx2             = _mm_add_pd(fjx2,tx);
1057             fjy2             = _mm_add_pd(fjy2,ty);
1058             fjz2             = _mm_add_pd(fjz2,tz);
1059
1060             }
1061
1062             /**************************
1063              * CALCULATE INTERACTIONS *
1064              **************************/
1065
1066             if (gmx_mm_any_lt(rsq23,rcutoff2))
1067             {
1068
1069             /* REACTION-FIELD ELECTROSTATICS */
1070             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
1071             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1072
1073             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1074
1075             /* Update potential sum for this i atom from the interaction with this j atom. */
1076             velec            = _mm_and_pd(velec,cutoff_mask);
1077             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1078             velecsum         = _mm_add_pd(velecsum,velec);
1079
1080             fscal            = felec;
1081
1082             fscal            = _mm_and_pd(fscal,cutoff_mask);
1083
1084             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1085
1086             /* Calculate temporary vectorial force */
1087             tx               = _mm_mul_pd(fscal,dx23);
1088             ty               = _mm_mul_pd(fscal,dy23);
1089             tz               = _mm_mul_pd(fscal,dz23);
1090
1091             /* Update vectorial force */
1092             fix2             = _mm_add_pd(fix2,tx);
1093             fiy2             = _mm_add_pd(fiy2,ty);
1094             fiz2             = _mm_add_pd(fiz2,tz);
1095
1096             fjx3             = _mm_add_pd(fjx3,tx);
1097             fjy3             = _mm_add_pd(fjy3,ty);
1098             fjz3             = _mm_add_pd(fjz3,tz);
1099
1100             }
1101
1102             /**************************
1103              * CALCULATE INTERACTIONS *
1104              **************************/
1105
1106             if (gmx_mm_any_lt(rsq31,rcutoff2))
1107             {
1108
1109             /* REACTION-FIELD ELECTROSTATICS */
1110             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
1111             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1112
1113             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1114
1115             /* Update potential sum for this i atom from the interaction with this j atom. */
1116             velec            = _mm_and_pd(velec,cutoff_mask);
1117             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1118             velecsum         = _mm_add_pd(velecsum,velec);
1119
1120             fscal            = felec;
1121
1122             fscal            = _mm_and_pd(fscal,cutoff_mask);
1123
1124             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1125
1126             /* Calculate temporary vectorial force */
1127             tx               = _mm_mul_pd(fscal,dx31);
1128             ty               = _mm_mul_pd(fscal,dy31);
1129             tz               = _mm_mul_pd(fscal,dz31);
1130
1131             /* Update vectorial force */
1132             fix3             = _mm_add_pd(fix3,tx);
1133             fiy3             = _mm_add_pd(fiy3,ty);
1134             fiz3             = _mm_add_pd(fiz3,tz);
1135
1136             fjx1             = _mm_add_pd(fjx1,tx);
1137             fjy1             = _mm_add_pd(fjy1,ty);
1138             fjz1             = _mm_add_pd(fjz1,tz);
1139
1140             }
1141
1142             /**************************
1143              * CALCULATE INTERACTIONS *
1144              **************************/
1145
1146             if (gmx_mm_any_lt(rsq32,rcutoff2))
1147             {
1148
1149             /* REACTION-FIELD ELECTROSTATICS */
1150             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
1151             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1152
1153             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1154
1155             /* Update potential sum for this i atom from the interaction with this j atom. */
1156             velec            = _mm_and_pd(velec,cutoff_mask);
1157             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1158             velecsum         = _mm_add_pd(velecsum,velec);
1159
1160             fscal            = felec;
1161
1162             fscal            = _mm_and_pd(fscal,cutoff_mask);
1163
1164             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1165
1166             /* Calculate temporary vectorial force */
1167             tx               = _mm_mul_pd(fscal,dx32);
1168             ty               = _mm_mul_pd(fscal,dy32);
1169             tz               = _mm_mul_pd(fscal,dz32);
1170
1171             /* Update vectorial force */
1172             fix3             = _mm_add_pd(fix3,tx);
1173             fiy3             = _mm_add_pd(fiy3,ty);
1174             fiz3             = _mm_add_pd(fiz3,tz);
1175
1176             fjx2             = _mm_add_pd(fjx2,tx);
1177             fjy2             = _mm_add_pd(fjy2,ty);
1178             fjz2             = _mm_add_pd(fjz2,tz);
1179
1180             }
1181
1182             /**************************
1183              * CALCULATE INTERACTIONS *
1184              **************************/
1185
1186             if (gmx_mm_any_lt(rsq33,rcutoff2))
1187             {
1188
1189             /* REACTION-FIELD ELECTROSTATICS */
1190             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
1191             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1192
1193             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1194
1195             /* Update potential sum for this i atom from the interaction with this j atom. */
1196             velec            = _mm_and_pd(velec,cutoff_mask);
1197             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1198             velecsum         = _mm_add_pd(velecsum,velec);
1199
1200             fscal            = felec;
1201
1202             fscal            = _mm_and_pd(fscal,cutoff_mask);
1203
1204             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1205
1206             /* Calculate temporary vectorial force */
1207             tx               = _mm_mul_pd(fscal,dx33);
1208             ty               = _mm_mul_pd(fscal,dy33);
1209             tz               = _mm_mul_pd(fscal,dz33);
1210
1211             /* Update vectorial force */
1212             fix3             = _mm_add_pd(fix3,tx);
1213             fiy3             = _mm_add_pd(fiy3,ty);
1214             fiz3             = _mm_add_pd(fiz3,tz);
1215
1216             fjx3             = _mm_add_pd(fjx3,tx);
1217             fjy3             = _mm_add_pd(fjy3,ty);
1218             fjz3             = _mm_add_pd(fjz3,tz);
1219
1220             }
1221
1222             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1223
1224             /* Inner loop uses 383 flops */
1225         }
1226
1227         /* End of innermost loop */
1228
1229         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1230                                               f+i_coord_offset,fshift+i_shift_offset);
1231
1232         ggid                        = gid[iidx];
1233         /* Update potential energies */
1234         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1235         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
1236
1237         /* Increment number of inner iterations */
1238         inneriter                  += j_index_end - j_index_start;
1239
1240         /* Outer loop uses 26 flops */
1241     }
1242
1243     /* Increment number of outer iterations */
1244     outeriter        += nri;
1245
1246     /* Update outer/inner flops */
1247
1248     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*383);
1249 }
1250 /*
1251  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW4W4_F_sse2_double
1252  * Electrostatics interaction: ReactionField
1253  * VdW interaction:            CubicSplineTable
1254  * Geometry:                   Water4-Water4
1255  * Calculate force/pot:        Force
1256  */
1257 void
1258 nb_kernel_ElecRFCut_VdwCSTab_GeomW4W4_F_sse2_double
1259                     (t_nblist                    * gmx_restrict       nlist,
1260                      rvec                        * gmx_restrict          xx,
1261                      rvec                        * gmx_restrict          ff,
1262                      struct t_forcerec           * gmx_restrict          fr,
1263                      t_mdatoms                   * gmx_restrict     mdatoms,
1264                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1265                      t_nrnb                      * gmx_restrict        nrnb)
1266 {
1267     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1268      * just 0 for non-waters.
1269      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1270      * jnr indices corresponding to data put in the four positions in the SIMD register.
1271      */
1272     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1273     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1274     int              jnrA,jnrB;
1275     int              j_coord_offsetA,j_coord_offsetB;
1276     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1277     real             rcutoff_scalar;
1278     real             *shiftvec,*fshift,*x,*f;
1279     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1280     int              vdwioffset0;
1281     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1282     int              vdwioffset1;
1283     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1284     int              vdwioffset2;
1285     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1286     int              vdwioffset3;
1287     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1288     int              vdwjidx0A,vdwjidx0B;
1289     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1290     int              vdwjidx1A,vdwjidx1B;
1291     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1292     int              vdwjidx2A,vdwjidx2B;
1293     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1294     int              vdwjidx3A,vdwjidx3B;
1295     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1296     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1297     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1298     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1299     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1300     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1301     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1302     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1303     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1304     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1305     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1306     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1307     real             *charge;
1308     int              nvdwtype;
1309     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1310     int              *vdwtype;
1311     real             *vdwparam;
1312     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1313     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1314     __m128i          vfitab;
1315     __m128i          ifour       = _mm_set1_epi32(4);
1316     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
1317     real             *vftab;
1318     __m128d          dummy_mask,cutoff_mask;
1319     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1320     __m128d          one     = _mm_set1_pd(1.0);
1321     __m128d          two     = _mm_set1_pd(2.0);
1322     x                = xx[0];
1323     f                = ff[0];
1324
1325     nri              = nlist->nri;
1326     iinr             = nlist->iinr;
1327     jindex           = nlist->jindex;
1328     jjnr             = nlist->jjnr;
1329     shiftidx         = nlist->shift;
1330     gid              = nlist->gid;
1331     shiftvec         = fr->shift_vec[0];
1332     fshift           = fr->fshift[0];
1333     facel            = _mm_set1_pd(fr->ic->epsfac);
1334     charge           = mdatoms->chargeA;
1335     krf              = _mm_set1_pd(fr->ic->k_rf);
1336     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1337     crf              = _mm_set1_pd(fr->ic->c_rf);
1338     nvdwtype         = fr->ntype;
1339     vdwparam         = fr->nbfp;
1340     vdwtype          = mdatoms->typeA;
1341
1342     vftab            = kernel_data->table_vdw->data;
1343     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
1344
1345     /* Setup water-specific parameters */
1346     inr              = nlist->iinr[0];
1347     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1348     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1349     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1350     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1351
1352     jq1              = _mm_set1_pd(charge[inr+1]);
1353     jq2              = _mm_set1_pd(charge[inr+2]);
1354     jq3              = _mm_set1_pd(charge[inr+3]);
1355     vdwjidx0A        = 2*vdwtype[inr+0];
1356     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1357     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1358     qq11             = _mm_mul_pd(iq1,jq1);
1359     qq12             = _mm_mul_pd(iq1,jq2);
1360     qq13             = _mm_mul_pd(iq1,jq3);
1361     qq21             = _mm_mul_pd(iq2,jq1);
1362     qq22             = _mm_mul_pd(iq2,jq2);
1363     qq23             = _mm_mul_pd(iq2,jq3);
1364     qq31             = _mm_mul_pd(iq3,jq1);
1365     qq32             = _mm_mul_pd(iq3,jq2);
1366     qq33             = _mm_mul_pd(iq3,jq3);
1367
1368     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1369     rcutoff_scalar   = fr->ic->rcoulomb;
1370     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1371     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1372
1373     /* Avoid stupid compiler warnings */
1374     jnrA = jnrB = 0;
1375     j_coord_offsetA = 0;
1376     j_coord_offsetB = 0;
1377
1378     outeriter        = 0;
1379     inneriter        = 0;
1380
1381     /* Start outer loop over neighborlists */
1382     for(iidx=0; iidx<nri; iidx++)
1383     {
1384         /* Load shift vector for this list */
1385         i_shift_offset   = DIM*shiftidx[iidx];
1386
1387         /* Load limits for loop over neighbors */
1388         j_index_start    = jindex[iidx];
1389         j_index_end      = jindex[iidx+1];
1390
1391         /* Get outer coordinate index */
1392         inr              = iinr[iidx];
1393         i_coord_offset   = DIM*inr;
1394
1395         /* Load i particle coords and add shift vector */
1396         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1397                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1398
1399         fix0             = _mm_setzero_pd();
1400         fiy0             = _mm_setzero_pd();
1401         fiz0             = _mm_setzero_pd();
1402         fix1             = _mm_setzero_pd();
1403         fiy1             = _mm_setzero_pd();
1404         fiz1             = _mm_setzero_pd();
1405         fix2             = _mm_setzero_pd();
1406         fiy2             = _mm_setzero_pd();
1407         fiz2             = _mm_setzero_pd();
1408         fix3             = _mm_setzero_pd();
1409         fiy3             = _mm_setzero_pd();
1410         fiz3             = _mm_setzero_pd();
1411
1412         /* Start inner kernel loop */
1413         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1414         {
1415
1416             /* Get j neighbor index, and coordinate index */
1417             jnrA             = jjnr[jidx];
1418             jnrB             = jjnr[jidx+1];
1419             j_coord_offsetA  = DIM*jnrA;
1420             j_coord_offsetB  = DIM*jnrB;
1421
1422             /* load j atom coordinates */
1423             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1424                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1425                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1426
1427             /* Calculate displacement vector */
1428             dx00             = _mm_sub_pd(ix0,jx0);
1429             dy00             = _mm_sub_pd(iy0,jy0);
1430             dz00             = _mm_sub_pd(iz0,jz0);
1431             dx11             = _mm_sub_pd(ix1,jx1);
1432             dy11             = _mm_sub_pd(iy1,jy1);
1433             dz11             = _mm_sub_pd(iz1,jz1);
1434             dx12             = _mm_sub_pd(ix1,jx2);
1435             dy12             = _mm_sub_pd(iy1,jy2);
1436             dz12             = _mm_sub_pd(iz1,jz2);
1437             dx13             = _mm_sub_pd(ix1,jx3);
1438             dy13             = _mm_sub_pd(iy1,jy3);
1439             dz13             = _mm_sub_pd(iz1,jz3);
1440             dx21             = _mm_sub_pd(ix2,jx1);
1441             dy21             = _mm_sub_pd(iy2,jy1);
1442             dz21             = _mm_sub_pd(iz2,jz1);
1443             dx22             = _mm_sub_pd(ix2,jx2);
1444             dy22             = _mm_sub_pd(iy2,jy2);
1445             dz22             = _mm_sub_pd(iz2,jz2);
1446             dx23             = _mm_sub_pd(ix2,jx3);
1447             dy23             = _mm_sub_pd(iy2,jy3);
1448             dz23             = _mm_sub_pd(iz2,jz3);
1449             dx31             = _mm_sub_pd(ix3,jx1);
1450             dy31             = _mm_sub_pd(iy3,jy1);
1451             dz31             = _mm_sub_pd(iz3,jz1);
1452             dx32             = _mm_sub_pd(ix3,jx2);
1453             dy32             = _mm_sub_pd(iy3,jy2);
1454             dz32             = _mm_sub_pd(iz3,jz2);
1455             dx33             = _mm_sub_pd(ix3,jx3);
1456             dy33             = _mm_sub_pd(iy3,jy3);
1457             dz33             = _mm_sub_pd(iz3,jz3);
1458
1459             /* Calculate squared distance and things based on it */
1460             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1461             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1462             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1463             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1464             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1465             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1466             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1467             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1468             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1469             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1470
1471             rinv00           = sse2_invsqrt_d(rsq00);
1472             rinv11           = sse2_invsqrt_d(rsq11);
1473             rinv12           = sse2_invsqrt_d(rsq12);
1474             rinv13           = sse2_invsqrt_d(rsq13);
1475             rinv21           = sse2_invsqrt_d(rsq21);
1476             rinv22           = sse2_invsqrt_d(rsq22);
1477             rinv23           = sse2_invsqrt_d(rsq23);
1478             rinv31           = sse2_invsqrt_d(rsq31);
1479             rinv32           = sse2_invsqrt_d(rsq32);
1480             rinv33           = sse2_invsqrt_d(rsq33);
1481
1482             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1483             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1484             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1485             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1486             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1487             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1488             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1489             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1490             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1491
1492             fjx0             = _mm_setzero_pd();
1493             fjy0             = _mm_setzero_pd();
1494             fjz0             = _mm_setzero_pd();
1495             fjx1             = _mm_setzero_pd();
1496             fjy1             = _mm_setzero_pd();
1497             fjz1             = _mm_setzero_pd();
1498             fjx2             = _mm_setzero_pd();
1499             fjy2             = _mm_setzero_pd();
1500             fjz2             = _mm_setzero_pd();
1501             fjx3             = _mm_setzero_pd();
1502             fjy3             = _mm_setzero_pd();
1503             fjz3             = _mm_setzero_pd();
1504
1505             /**************************
1506              * CALCULATE INTERACTIONS *
1507              **************************/
1508
1509             r00              = _mm_mul_pd(rsq00,rinv00);
1510
1511             /* Calculate table index by multiplying r with table scale and truncate to integer */
1512             rt               = _mm_mul_pd(r00,vftabscale);
1513             vfitab           = _mm_cvttpd_epi32(rt);
1514             vfeps            = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
1515             vfitab           = _mm_slli_epi32(vfitab,3);
1516
1517             /* CUBIC SPLINE TABLE DISPERSION */
1518             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1519             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
1520             GMX_MM_TRANSPOSE2_PD(Y,F);
1521             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1522             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
1523             GMX_MM_TRANSPOSE2_PD(G,H);
1524             Heps             = _mm_mul_pd(vfeps,H);
1525             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1526             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1527             fvdw6            = _mm_mul_pd(c6_00,FF);
1528
1529             /* CUBIC SPLINE TABLE REPULSION */
1530             vfitab           = _mm_add_epi32(vfitab,ifour);
1531             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1532             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
1533             GMX_MM_TRANSPOSE2_PD(Y,F);
1534             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1535             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
1536             GMX_MM_TRANSPOSE2_PD(G,H);
1537             Heps             = _mm_mul_pd(vfeps,H);
1538             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1539             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1540             fvdw12           = _mm_mul_pd(c12_00,FF);
1541             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1542
1543             fscal            = fvdw;
1544
1545             /* Calculate temporary vectorial force */
1546             tx               = _mm_mul_pd(fscal,dx00);
1547             ty               = _mm_mul_pd(fscal,dy00);
1548             tz               = _mm_mul_pd(fscal,dz00);
1549
1550             /* Update vectorial force */
1551             fix0             = _mm_add_pd(fix0,tx);
1552             fiy0             = _mm_add_pd(fiy0,ty);
1553             fiz0             = _mm_add_pd(fiz0,tz);
1554
1555             fjx0             = _mm_add_pd(fjx0,tx);
1556             fjy0             = _mm_add_pd(fjy0,ty);
1557             fjz0             = _mm_add_pd(fjz0,tz);
1558
1559             /**************************
1560              * CALCULATE INTERACTIONS *
1561              **************************/
1562
1563             if (gmx_mm_any_lt(rsq11,rcutoff2))
1564             {
1565
1566             /* REACTION-FIELD ELECTROSTATICS */
1567             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1568
1569             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1570
1571             fscal            = felec;
1572
1573             fscal            = _mm_and_pd(fscal,cutoff_mask);
1574
1575             /* Calculate temporary vectorial force */
1576             tx               = _mm_mul_pd(fscal,dx11);
1577             ty               = _mm_mul_pd(fscal,dy11);
1578             tz               = _mm_mul_pd(fscal,dz11);
1579
1580             /* Update vectorial force */
1581             fix1             = _mm_add_pd(fix1,tx);
1582             fiy1             = _mm_add_pd(fiy1,ty);
1583             fiz1             = _mm_add_pd(fiz1,tz);
1584
1585             fjx1             = _mm_add_pd(fjx1,tx);
1586             fjy1             = _mm_add_pd(fjy1,ty);
1587             fjz1             = _mm_add_pd(fjz1,tz);
1588
1589             }
1590
1591             /**************************
1592              * CALCULATE INTERACTIONS *
1593              **************************/
1594
1595             if (gmx_mm_any_lt(rsq12,rcutoff2))
1596             {
1597
1598             /* REACTION-FIELD ELECTROSTATICS */
1599             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1600
1601             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1602
1603             fscal            = felec;
1604
1605             fscal            = _mm_and_pd(fscal,cutoff_mask);
1606
1607             /* Calculate temporary vectorial force */
1608             tx               = _mm_mul_pd(fscal,dx12);
1609             ty               = _mm_mul_pd(fscal,dy12);
1610             tz               = _mm_mul_pd(fscal,dz12);
1611
1612             /* Update vectorial force */
1613             fix1             = _mm_add_pd(fix1,tx);
1614             fiy1             = _mm_add_pd(fiy1,ty);
1615             fiz1             = _mm_add_pd(fiz1,tz);
1616
1617             fjx2             = _mm_add_pd(fjx2,tx);
1618             fjy2             = _mm_add_pd(fjy2,ty);
1619             fjz2             = _mm_add_pd(fjz2,tz);
1620
1621             }
1622
1623             /**************************
1624              * CALCULATE INTERACTIONS *
1625              **************************/
1626
1627             if (gmx_mm_any_lt(rsq13,rcutoff2))
1628             {
1629
1630             /* REACTION-FIELD ELECTROSTATICS */
1631             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
1632
1633             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1634
1635             fscal            = felec;
1636
1637             fscal            = _mm_and_pd(fscal,cutoff_mask);
1638
1639             /* Calculate temporary vectorial force */
1640             tx               = _mm_mul_pd(fscal,dx13);
1641             ty               = _mm_mul_pd(fscal,dy13);
1642             tz               = _mm_mul_pd(fscal,dz13);
1643
1644             /* Update vectorial force */
1645             fix1             = _mm_add_pd(fix1,tx);
1646             fiy1             = _mm_add_pd(fiy1,ty);
1647             fiz1             = _mm_add_pd(fiz1,tz);
1648
1649             fjx3             = _mm_add_pd(fjx3,tx);
1650             fjy3             = _mm_add_pd(fjy3,ty);
1651             fjz3             = _mm_add_pd(fjz3,tz);
1652
1653             }
1654
1655             /**************************
1656              * CALCULATE INTERACTIONS *
1657              **************************/
1658
1659             if (gmx_mm_any_lt(rsq21,rcutoff2))
1660             {
1661
1662             /* REACTION-FIELD ELECTROSTATICS */
1663             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1664
1665             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1666
1667             fscal            = felec;
1668
1669             fscal            = _mm_and_pd(fscal,cutoff_mask);
1670
1671             /* Calculate temporary vectorial force */
1672             tx               = _mm_mul_pd(fscal,dx21);
1673             ty               = _mm_mul_pd(fscal,dy21);
1674             tz               = _mm_mul_pd(fscal,dz21);
1675
1676             /* Update vectorial force */
1677             fix2             = _mm_add_pd(fix2,tx);
1678             fiy2             = _mm_add_pd(fiy2,ty);
1679             fiz2             = _mm_add_pd(fiz2,tz);
1680
1681             fjx1             = _mm_add_pd(fjx1,tx);
1682             fjy1             = _mm_add_pd(fjy1,ty);
1683             fjz1             = _mm_add_pd(fjz1,tz);
1684
1685             }
1686
1687             /**************************
1688              * CALCULATE INTERACTIONS *
1689              **************************/
1690
1691             if (gmx_mm_any_lt(rsq22,rcutoff2))
1692             {
1693
1694             /* REACTION-FIELD ELECTROSTATICS */
1695             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1696
1697             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1698
1699             fscal            = felec;
1700
1701             fscal            = _mm_and_pd(fscal,cutoff_mask);
1702
1703             /* Calculate temporary vectorial force */
1704             tx               = _mm_mul_pd(fscal,dx22);
1705             ty               = _mm_mul_pd(fscal,dy22);
1706             tz               = _mm_mul_pd(fscal,dz22);
1707
1708             /* Update vectorial force */
1709             fix2             = _mm_add_pd(fix2,tx);
1710             fiy2             = _mm_add_pd(fiy2,ty);
1711             fiz2             = _mm_add_pd(fiz2,tz);
1712
1713             fjx2             = _mm_add_pd(fjx2,tx);
1714             fjy2             = _mm_add_pd(fjy2,ty);
1715             fjz2             = _mm_add_pd(fjz2,tz);
1716
1717             }
1718
1719             /**************************
1720              * CALCULATE INTERACTIONS *
1721              **************************/
1722
1723             if (gmx_mm_any_lt(rsq23,rcutoff2))
1724             {
1725
1726             /* REACTION-FIELD ELECTROSTATICS */
1727             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1728
1729             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1730
1731             fscal            = felec;
1732
1733             fscal            = _mm_and_pd(fscal,cutoff_mask);
1734
1735             /* Calculate temporary vectorial force */
1736             tx               = _mm_mul_pd(fscal,dx23);
1737             ty               = _mm_mul_pd(fscal,dy23);
1738             tz               = _mm_mul_pd(fscal,dz23);
1739
1740             /* Update vectorial force */
1741             fix2             = _mm_add_pd(fix2,tx);
1742             fiy2             = _mm_add_pd(fiy2,ty);
1743             fiz2             = _mm_add_pd(fiz2,tz);
1744
1745             fjx3             = _mm_add_pd(fjx3,tx);
1746             fjy3             = _mm_add_pd(fjy3,ty);
1747             fjz3             = _mm_add_pd(fjz3,tz);
1748
1749             }
1750
1751             /**************************
1752              * CALCULATE INTERACTIONS *
1753              **************************/
1754
1755             if (gmx_mm_any_lt(rsq31,rcutoff2))
1756             {
1757
1758             /* REACTION-FIELD ELECTROSTATICS */
1759             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1760
1761             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1762
1763             fscal            = felec;
1764
1765             fscal            = _mm_and_pd(fscal,cutoff_mask);
1766
1767             /* Calculate temporary vectorial force */
1768             tx               = _mm_mul_pd(fscal,dx31);
1769             ty               = _mm_mul_pd(fscal,dy31);
1770             tz               = _mm_mul_pd(fscal,dz31);
1771
1772             /* Update vectorial force */
1773             fix3             = _mm_add_pd(fix3,tx);
1774             fiy3             = _mm_add_pd(fiy3,ty);
1775             fiz3             = _mm_add_pd(fiz3,tz);
1776
1777             fjx1             = _mm_add_pd(fjx1,tx);
1778             fjy1             = _mm_add_pd(fjy1,ty);
1779             fjz1             = _mm_add_pd(fjz1,tz);
1780
1781             }
1782
1783             /**************************
1784              * CALCULATE INTERACTIONS *
1785              **************************/
1786
1787             if (gmx_mm_any_lt(rsq32,rcutoff2))
1788             {
1789
1790             /* REACTION-FIELD ELECTROSTATICS */
1791             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1792
1793             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1794
1795             fscal            = felec;
1796
1797             fscal            = _mm_and_pd(fscal,cutoff_mask);
1798
1799             /* Calculate temporary vectorial force */
1800             tx               = _mm_mul_pd(fscal,dx32);
1801             ty               = _mm_mul_pd(fscal,dy32);
1802             tz               = _mm_mul_pd(fscal,dz32);
1803
1804             /* Update vectorial force */
1805             fix3             = _mm_add_pd(fix3,tx);
1806             fiy3             = _mm_add_pd(fiy3,ty);
1807             fiz3             = _mm_add_pd(fiz3,tz);
1808
1809             fjx2             = _mm_add_pd(fjx2,tx);
1810             fjy2             = _mm_add_pd(fjy2,ty);
1811             fjz2             = _mm_add_pd(fjz2,tz);
1812
1813             }
1814
1815             /**************************
1816              * CALCULATE INTERACTIONS *
1817              **************************/
1818
1819             if (gmx_mm_any_lt(rsq33,rcutoff2))
1820             {
1821
1822             /* REACTION-FIELD ELECTROSTATICS */
1823             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1824
1825             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1826
1827             fscal            = felec;
1828
1829             fscal            = _mm_and_pd(fscal,cutoff_mask);
1830
1831             /* Calculate temporary vectorial force */
1832             tx               = _mm_mul_pd(fscal,dx33);
1833             ty               = _mm_mul_pd(fscal,dy33);
1834             tz               = _mm_mul_pd(fscal,dz33);
1835
1836             /* Update vectorial force */
1837             fix3             = _mm_add_pd(fix3,tx);
1838             fiy3             = _mm_add_pd(fiy3,ty);
1839             fiz3             = _mm_add_pd(fiz3,tz);
1840
1841             fjx3             = _mm_add_pd(fjx3,tx);
1842             fjy3             = _mm_add_pd(fjy3,ty);
1843             fjz3             = _mm_add_pd(fjz3,tz);
1844
1845             }
1846
1847             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);
1848
1849             /* Inner loop uses 321 flops */
1850         }
1851
1852         if(jidx<j_index_end)
1853         {
1854
1855             jnrA             = jjnr[jidx];
1856             j_coord_offsetA  = DIM*jnrA;
1857
1858             /* load j atom coordinates */
1859             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1860                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1861                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1862
1863             /* Calculate displacement vector */
1864             dx00             = _mm_sub_pd(ix0,jx0);
1865             dy00             = _mm_sub_pd(iy0,jy0);
1866             dz00             = _mm_sub_pd(iz0,jz0);
1867             dx11             = _mm_sub_pd(ix1,jx1);
1868             dy11             = _mm_sub_pd(iy1,jy1);
1869             dz11             = _mm_sub_pd(iz1,jz1);
1870             dx12             = _mm_sub_pd(ix1,jx2);
1871             dy12             = _mm_sub_pd(iy1,jy2);
1872             dz12             = _mm_sub_pd(iz1,jz2);
1873             dx13             = _mm_sub_pd(ix1,jx3);
1874             dy13             = _mm_sub_pd(iy1,jy3);
1875             dz13             = _mm_sub_pd(iz1,jz3);
1876             dx21             = _mm_sub_pd(ix2,jx1);
1877             dy21             = _mm_sub_pd(iy2,jy1);
1878             dz21             = _mm_sub_pd(iz2,jz1);
1879             dx22             = _mm_sub_pd(ix2,jx2);
1880             dy22             = _mm_sub_pd(iy2,jy2);
1881             dz22             = _mm_sub_pd(iz2,jz2);
1882             dx23             = _mm_sub_pd(ix2,jx3);
1883             dy23             = _mm_sub_pd(iy2,jy3);
1884             dz23             = _mm_sub_pd(iz2,jz3);
1885             dx31             = _mm_sub_pd(ix3,jx1);
1886             dy31             = _mm_sub_pd(iy3,jy1);
1887             dz31             = _mm_sub_pd(iz3,jz1);
1888             dx32             = _mm_sub_pd(ix3,jx2);
1889             dy32             = _mm_sub_pd(iy3,jy2);
1890             dz32             = _mm_sub_pd(iz3,jz2);
1891             dx33             = _mm_sub_pd(ix3,jx3);
1892             dy33             = _mm_sub_pd(iy3,jy3);
1893             dz33             = _mm_sub_pd(iz3,jz3);
1894
1895             /* Calculate squared distance and things based on it */
1896             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1897             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1898             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1899             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1900             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1901             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1902             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1903             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1904             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1905             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1906
1907             rinv00           = sse2_invsqrt_d(rsq00);
1908             rinv11           = sse2_invsqrt_d(rsq11);
1909             rinv12           = sse2_invsqrt_d(rsq12);
1910             rinv13           = sse2_invsqrt_d(rsq13);
1911             rinv21           = sse2_invsqrt_d(rsq21);
1912             rinv22           = sse2_invsqrt_d(rsq22);
1913             rinv23           = sse2_invsqrt_d(rsq23);
1914             rinv31           = sse2_invsqrt_d(rsq31);
1915             rinv32           = sse2_invsqrt_d(rsq32);
1916             rinv33           = sse2_invsqrt_d(rsq33);
1917
1918             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1919             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1920             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1921             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1922             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1923             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1924             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1925             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1926             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1927
1928             fjx0             = _mm_setzero_pd();
1929             fjy0             = _mm_setzero_pd();
1930             fjz0             = _mm_setzero_pd();
1931             fjx1             = _mm_setzero_pd();
1932             fjy1             = _mm_setzero_pd();
1933             fjz1             = _mm_setzero_pd();
1934             fjx2             = _mm_setzero_pd();
1935             fjy2             = _mm_setzero_pd();
1936             fjz2             = _mm_setzero_pd();
1937             fjx3             = _mm_setzero_pd();
1938             fjy3             = _mm_setzero_pd();
1939             fjz3             = _mm_setzero_pd();
1940
1941             /**************************
1942              * CALCULATE INTERACTIONS *
1943              **************************/
1944
1945             r00              = _mm_mul_pd(rsq00,rinv00);
1946
1947             /* Calculate table index by multiplying r with table scale and truncate to integer */
1948             rt               = _mm_mul_pd(r00,vftabscale);
1949             vfitab           = _mm_cvttpd_epi32(rt);
1950             vfeps            = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
1951             vfitab           = _mm_slli_epi32(vfitab,3);
1952
1953             /* CUBIC SPLINE TABLE DISPERSION */
1954             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1955             F                = _mm_setzero_pd();
1956             GMX_MM_TRANSPOSE2_PD(Y,F);
1957             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1958             H                = _mm_setzero_pd();
1959             GMX_MM_TRANSPOSE2_PD(G,H);
1960             Heps             = _mm_mul_pd(vfeps,H);
1961             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1962             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1963             fvdw6            = _mm_mul_pd(c6_00,FF);
1964
1965             /* CUBIC SPLINE TABLE REPULSION */
1966             vfitab           = _mm_add_epi32(vfitab,ifour);
1967             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1968             F                = _mm_setzero_pd();
1969             GMX_MM_TRANSPOSE2_PD(Y,F);
1970             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1971             H                = _mm_setzero_pd();
1972             GMX_MM_TRANSPOSE2_PD(G,H);
1973             Heps             = _mm_mul_pd(vfeps,H);
1974             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1975             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1976             fvdw12           = _mm_mul_pd(c12_00,FF);
1977             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1978
1979             fscal            = fvdw;
1980
1981             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1982
1983             /* Calculate temporary vectorial force */
1984             tx               = _mm_mul_pd(fscal,dx00);
1985             ty               = _mm_mul_pd(fscal,dy00);
1986             tz               = _mm_mul_pd(fscal,dz00);
1987
1988             /* Update vectorial force */
1989             fix0             = _mm_add_pd(fix0,tx);
1990             fiy0             = _mm_add_pd(fiy0,ty);
1991             fiz0             = _mm_add_pd(fiz0,tz);
1992
1993             fjx0             = _mm_add_pd(fjx0,tx);
1994             fjy0             = _mm_add_pd(fjy0,ty);
1995             fjz0             = _mm_add_pd(fjz0,tz);
1996
1997             /**************************
1998              * CALCULATE INTERACTIONS *
1999              **************************/
2000
2001             if (gmx_mm_any_lt(rsq11,rcutoff2))
2002             {
2003
2004             /* REACTION-FIELD ELECTROSTATICS */
2005             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
2006
2007             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
2008
2009             fscal            = felec;
2010
2011             fscal            = _mm_and_pd(fscal,cutoff_mask);
2012
2013             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2014
2015             /* Calculate temporary vectorial force */
2016             tx               = _mm_mul_pd(fscal,dx11);
2017             ty               = _mm_mul_pd(fscal,dy11);
2018             tz               = _mm_mul_pd(fscal,dz11);
2019
2020             /* Update vectorial force */
2021             fix1             = _mm_add_pd(fix1,tx);
2022             fiy1             = _mm_add_pd(fiy1,ty);
2023             fiz1             = _mm_add_pd(fiz1,tz);
2024
2025             fjx1             = _mm_add_pd(fjx1,tx);
2026             fjy1             = _mm_add_pd(fjy1,ty);
2027             fjz1             = _mm_add_pd(fjz1,tz);
2028
2029             }
2030
2031             /**************************
2032              * CALCULATE INTERACTIONS *
2033              **************************/
2034
2035             if (gmx_mm_any_lt(rsq12,rcutoff2))
2036             {
2037
2038             /* REACTION-FIELD ELECTROSTATICS */
2039             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
2040
2041             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
2042
2043             fscal            = felec;
2044
2045             fscal            = _mm_and_pd(fscal,cutoff_mask);
2046
2047             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2048
2049             /* Calculate temporary vectorial force */
2050             tx               = _mm_mul_pd(fscal,dx12);
2051             ty               = _mm_mul_pd(fscal,dy12);
2052             tz               = _mm_mul_pd(fscal,dz12);
2053
2054             /* Update vectorial force */
2055             fix1             = _mm_add_pd(fix1,tx);
2056             fiy1             = _mm_add_pd(fiy1,ty);
2057             fiz1             = _mm_add_pd(fiz1,tz);
2058
2059             fjx2             = _mm_add_pd(fjx2,tx);
2060             fjy2             = _mm_add_pd(fjy2,ty);
2061             fjz2             = _mm_add_pd(fjz2,tz);
2062
2063             }
2064
2065             /**************************
2066              * CALCULATE INTERACTIONS *
2067              **************************/
2068
2069             if (gmx_mm_any_lt(rsq13,rcutoff2))
2070             {
2071
2072             /* REACTION-FIELD ELECTROSTATICS */
2073             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
2074
2075             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
2076
2077             fscal            = felec;
2078
2079             fscal            = _mm_and_pd(fscal,cutoff_mask);
2080
2081             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2082
2083             /* Calculate temporary vectorial force */
2084             tx               = _mm_mul_pd(fscal,dx13);
2085             ty               = _mm_mul_pd(fscal,dy13);
2086             tz               = _mm_mul_pd(fscal,dz13);
2087
2088             /* Update vectorial force */
2089             fix1             = _mm_add_pd(fix1,tx);
2090             fiy1             = _mm_add_pd(fiy1,ty);
2091             fiz1             = _mm_add_pd(fiz1,tz);
2092
2093             fjx3             = _mm_add_pd(fjx3,tx);
2094             fjy3             = _mm_add_pd(fjy3,ty);
2095             fjz3             = _mm_add_pd(fjz3,tz);
2096
2097             }
2098
2099             /**************************
2100              * CALCULATE INTERACTIONS *
2101              **************************/
2102
2103             if (gmx_mm_any_lt(rsq21,rcutoff2))
2104             {
2105
2106             /* REACTION-FIELD ELECTROSTATICS */
2107             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
2108
2109             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
2110
2111             fscal            = felec;
2112
2113             fscal            = _mm_and_pd(fscal,cutoff_mask);
2114
2115             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2116
2117             /* Calculate temporary vectorial force */
2118             tx               = _mm_mul_pd(fscal,dx21);
2119             ty               = _mm_mul_pd(fscal,dy21);
2120             tz               = _mm_mul_pd(fscal,dz21);
2121
2122             /* Update vectorial force */
2123             fix2             = _mm_add_pd(fix2,tx);
2124             fiy2             = _mm_add_pd(fiy2,ty);
2125             fiz2             = _mm_add_pd(fiz2,tz);
2126
2127             fjx1             = _mm_add_pd(fjx1,tx);
2128             fjy1             = _mm_add_pd(fjy1,ty);
2129             fjz1             = _mm_add_pd(fjz1,tz);
2130
2131             }
2132
2133             /**************************
2134              * CALCULATE INTERACTIONS *
2135              **************************/
2136
2137             if (gmx_mm_any_lt(rsq22,rcutoff2))
2138             {
2139
2140             /* REACTION-FIELD ELECTROSTATICS */
2141             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
2142
2143             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
2144
2145             fscal            = felec;
2146
2147             fscal            = _mm_and_pd(fscal,cutoff_mask);
2148
2149             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2150
2151             /* Calculate temporary vectorial force */
2152             tx               = _mm_mul_pd(fscal,dx22);
2153             ty               = _mm_mul_pd(fscal,dy22);
2154             tz               = _mm_mul_pd(fscal,dz22);
2155
2156             /* Update vectorial force */
2157             fix2             = _mm_add_pd(fix2,tx);
2158             fiy2             = _mm_add_pd(fiy2,ty);
2159             fiz2             = _mm_add_pd(fiz2,tz);
2160
2161             fjx2             = _mm_add_pd(fjx2,tx);
2162             fjy2             = _mm_add_pd(fjy2,ty);
2163             fjz2             = _mm_add_pd(fjz2,tz);
2164
2165             }
2166
2167             /**************************
2168              * CALCULATE INTERACTIONS *
2169              **************************/
2170
2171             if (gmx_mm_any_lt(rsq23,rcutoff2))
2172             {
2173
2174             /* REACTION-FIELD ELECTROSTATICS */
2175             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
2176
2177             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
2178
2179             fscal            = felec;
2180
2181             fscal            = _mm_and_pd(fscal,cutoff_mask);
2182
2183             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2184
2185             /* Calculate temporary vectorial force */
2186             tx               = _mm_mul_pd(fscal,dx23);
2187             ty               = _mm_mul_pd(fscal,dy23);
2188             tz               = _mm_mul_pd(fscal,dz23);
2189
2190             /* Update vectorial force */
2191             fix2             = _mm_add_pd(fix2,tx);
2192             fiy2             = _mm_add_pd(fiy2,ty);
2193             fiz2             = _mm_add_pd(fiz2,tz);
2194
2195             fjx3             = _mm_add_pd(fjx3,tx);
2196             fjy3             = _mm_add_pd(fjy3,ty);
2197             fjz3             = _mm_add_pd(fjz3,tz);
2198
2199             }
2200
2201             /**************************
2202              * CALCULATE INTERACTIONS *
2203              **************************/
2204
2205             if (gmx_mm_any_lt(rsq31,rcutoff2))
2206             {
2207
2208             /* REACTION-FIELD ELECTROSTATICS */
2209             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
2210
2211             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
2212
2213             fscal            = felec;
2214
2215             fscal            = _mm_and_pd(fscal,cutoff_mask);
2216
2217             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2218
2219             /* Calculate temporary vectorial force */
2220             tx               = _mm_mul_pd(fscal,dx31);
2221             ty               = _mm_mul_pd(fscal,dy31);
2222             tz               = _mm_mul_pd(fscal,dz31);
2223
2224             /* Update vectorial force */
2225             fix3             = _mm_add_pd(fix3,tx);
2226             fiy3             = _mm_add_pd(fiy3,ty);
2227             fiz3             = _mm_add_pd(fiz3,tz);
2228
2229             fjx1             = _mm_add_pd(fjx1,tx);
2230             fjy1             = _mm_add_pd(fjy1,ty);
2231             fjz1             = _mm_add_pd(fjz1,tz);
2232
2233             }
2234
2235             /**************************
2236              * CALCULATE INTERACTIONS *
2237              **************************/
2238
2239             if (gmx_mm_any_lt(rsq32,rcutoff2))
2240             {
2241
2242             /* REACTION-FIELD ELECTROSTATICS */
2243             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
2244
2245             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
2246
2247             fscal            = felec;
2248
2249             fscal            = _mm_and_pd(fscal,cutoff_mask);
2250
2251             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2252
2253             /* Calculate temporary vectorial force */
2254             tx               = _mm_mul_pd(fscal,dx32);
2255             ty               = _mm_mul_pd(fscal,dy32);
2256             tz               = _mm_mul_pd(fscal,dz32);
2257
2258             /* Update vectorial force */
2259             fix3             = _mm_add_pd(fix3,tx);
2260             fiy3             = _mm_add_pd(fiy3,ty);
2261             fiz3             = _mm_add_pd(fiz3,tz);
2262
2263             fjx2             = _mm_add_pd(fjx2,tx);
2264             fjy2             = _mm_add_pd(fjy2,ty);
2265             fjz2             = _mm_add_pd(fjz2,tz);
2266
2267             }
2268
2269             /**************************
2270              * CALCULATE INTERACTIONS *
2271              **************************/
2272
2273             if (gmx_mm_any_lt(rsq33,rcutoff2))
2274             {
2275
2276             /* REACTION-FIELD ELECTROSTATICS */
2277             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
2278
2279             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
2280
2281             fscal            = felec;
2282
2283             fscal            = _mm_and_pd(fscal,cutoff_mask);
2284
2285             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2286
2287             /* Calculate temporary vectorial force */
2288             tx               = _mm_mul_pd(fscal,dx33);
2289             ty               = _mm_mul_pd(fscal,dy33);
2290             tz               = _mm_mul_pd(fscal,dz33);
2291
2292             /* Update vectorial force */
2293             fix3             = _mm_add_pd(fix3,tx);
2294             fiy3             = _mm_add_pd(fiy3,ty);
2295             fiz3             = _mm_add_pd(fiz3,tz);
2296
2297             fjx3             = _mm_add_pd(fjx3,tx);
2298             fjy3             = _mm_add_pd(fjy3,ty);
2299             fjz3             = _mm_add_pd(fjz3,tz);
2300
2301             }
2302
2303             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2304
2305             /* Inner loop uses 321 flops */
2306         }
2307
2308         /* End of innermost loop */
2309
2310         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2311                                               f+i_coord_offset,fshift+i_shift_offset);
2312
2313         /* Increment number of inner iterations */
2314         inneriter                  += j_index_end - j_index_start;
2315
2316         /* Outer loop uses 24 flops */
2317     }
2318
2319     /* Increment number of outer iterations */
2320     outeriter        += nri;
2321
2322     /* Update outer/inner flops */
2323
2324     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*321);
2325 }