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, 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/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_sse2_double.h"
  50 #include "kernelutil_x86_sse2_double.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW4W4_VF_sse2_double
  54  * Electrostatics interaction: ReactionField
  55  * VdW interaction:            CubicSplineTable
  56  * Geometry:                   Water4-Water4
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecRFCut_VdwCSTab_GeomW4W4_VF_sse2_double
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB;
  77     int              j_coord_offsetA,j_coord_offsetB;
  78     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  79     real             rcutoff_scalar;
  80     real             *shiftvec,*fshift,*x,*f;
  81     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwioffset1;
  85     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  86     int              vdwioffset2;
  87     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  88     int              vdwioffset3;
  89     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  90     int              vdwjidx0A,vdwjidx0B;
  91     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  92     int              vdwjidx1A,vdwjidx1B;
  93     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  94     int              vdwjidx2A,vdwjidx2B;
  95     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  96     int              vdwjidx3A,vdwjidx3B;
  97     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  98     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  99     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 100     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 101     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 102     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 103     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 104     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 105     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 106     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 107     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 108     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 109     real             *charge;
 110     int              nvdwtype;
 111     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 112     int              *vdwtype;
 113     real             *vdwparam;
 114     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 115     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 116     __m128i          vfitab;
 117     __m128i          ifour       = _mm_set1_epi32(4);
 118     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 119     real             *vftab;
 120     __m128d          dummy_mask,cutoff_mask;
 121     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 122     __m128d          one     = _mm_set1_pd(1.0);
 123     __m128d          two     = _mm_set1_pd(2.0);
 124     x                = xx[0];
 125     f                = ff[0];
 126
 127     nri              = nlist->nri;
 128     iinr             = nlist->iinr;
 129     jindex           = nlist->jindex;
 130     jjnr             = nlist->jjnr;
 131     shiftidx         = nlist->shift;
 132     gid              = nlist->gid;
 133     shiftvec         = fr->shift_vec[0];
 134     fshift           = fr->fshift[0];
 135     facel            = _mm_set1_pd(fr->epsfac);
 136     charge           = mdatoms->chargeA;
 137     krf              = _mm_set1_pd(fr->ic->k_rf);
 138     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 139     crf              = _mm_set1_pd(fr->ic->c_rf);
 140     nvdwtype         = fr->ntype;
 141     vdwparam         = fr->nbfp;
 142     vdwtype          = mdatoms->typeA;
 143
 144     vftab            = kernel_data->table_vdw->data;
 145     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
 146
 147     /* Setup water-specific parameters */
 148     inr              = nlist->iinr[0];
 149     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 150     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 151     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 152     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 153
 154     jq1              = _mm_set1_pd(charge[inr+1]);
 155     jq2              = _mm_set1_pd(charge[inr+2]);
 156     jq3              = _mm_set1_pd(charge[inr+3]);
 157     vdwjidx0A        = 2*vdwtype[inr+0];
 158     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 159     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 160     qq11             = _mm_mul_pd(iq1,jq1);
 161     qq12             = _mm_mul_pd(iq1,jq2);
 162     qq13             = _mm_mul_pd(iq1,jq3);
 163     qq21             = _mm_mul_pd(iq2,jq1);
 164     qq22             = _mm_mul_pd(iq2,jq2);
 165     qq23             = _mm_mul_pd(iq2,jq3);
 166     qq31             = _mm_mul_pd(iq3,jq1);
 167     qq32             = _mm_mul_pd(iq3,jq2);
 168     qq33             = _mm_mul_pd(iq3,jq3);
 169
 170     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 171     rcutoff_scalar   = fr->rcoulomb;
 172     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 173     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 174
 175     /* Avoid stupid compiler warnings */
 176     jnrA = jnrB = 0;
 177     j_coord_offsetA = 0;
 178     j_coord_offsetB = 0;
 179
 180     outeriter        = 0;
 181     inneriter        = 0;
 182
 183     /* Start outer loop over neighborlists */
 184     for(iidx=0; iidx<nri; iidx++)
 185     {
 186         /* Load shift vector for this list */
 187         i_shift_offset   = DIM*shiftidx[iidx];
 188
 189         /* Load limits for loop over neighbors */
 190         j_index_start    = jindex[iidx];
 191         j_index_end      = jindex[iidx+1];
 192
 193         /* Get outer coordinate index */
 194         inr              = iinr[iidx];
 195         i_coord_offset   = DIM*inr;
 196
 197         /* Load i particle coords and add shift vector */
 198         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 199                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 200
 201         fix0             = _mm_setzero_pd();
 202         fiy0             = _mm_setzero_pd();
 203         fiz0             = _mm_setzero_pd();
 204         fix1             = _mm_setzero_pd();
 205         fiy1             = _mm_setzero_pd();
 206         fiz1             = _mm_setzero_pd();
 207         fix2             = _mm_setzero_pd();
 208         fiy2             = _mm_setzero_pd();
 209         fiz2             = _mm_setzero_pd();
 210         fix3             = _mm_setzero_pd();
 211         fiy3             = _mm_setzero_pd();
 212         fiz3             = _mm_setzero_pd();
 213
 214         /* Reset potential sums */
 215         velecsum         = _mm_setzero_pd();
 216         vvdwsum          = _mm_setzero_pd();
 217
 218         /* Start inner kernel loop */
 219         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 220         {
 221
 222             /* Get j neighbor index, and coordinate index */
 223             jnrA             = jjnr[jidx];
 224             jnrB             = jjnr[jidx+1];
 225             j_coord_offsetA  = DIM*jnrA;
 226             j_coord_offsetB  = DIM*jnrB;
 227
 228             /* load j atom coordinates */
 229             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 230                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 231                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 232
 233             /* Calculate displacement vector */
 234             dx00             = _mm_sub_pd(ix0,jx0);
 235             dy00             = _mm_sub_pd(iy0,jy0);
 236             dz00             = _mm_sub_pd(iz0,jz0);
 237             dx11             = _mm_sub_pd(ix1,jx1);
 238             dy11             = _mm_sub_pd(iy1,jy1);
 239             dz11             = _mm_sub_pd(iz1,jz1);
 240             dx12             = _mm_sub_pd(ix1,jx2);
 241             dy12             = _mm_sub_pd(iy1,jy2);
 242             dz12             = _mm_sub_pd(iz1,jz2);
 243             dx13             = _mm_sub_pd(ix1,jx3);
 244             dy13             = _mm_sub_pd(iy1,jy3);
 245             dz13             = _mm_sub_pd(iz1,jz3);
 246             dx21             = _mm_sub_pd(ix2,jx1);
 247             dy21             = _mm_sub_pd(iy2,jy1);
 248             dz21             = _mm_sub_pd(iz2,jz1);
 249             dx22             = _mm_sub_pd(ix2,jx2);
 250             dy22             = _mm_sub_pd(iy2,jy2);
 251             dz22             = _mm_sub_pd(iz2,jz2);
 252             dx23             = _mm_sub_pd(ix2,jx3);
 253             dy23             = _mm_sub_pd(iy2,jy3);
 254             dz23             = _mm_sub_pd(iz2,jz3);
 255             dx31             = _mm_sub_pd(ix3,jx1);
 256             dy31             = _mm_sub_pd(iy3,jy1);
 257             dz31             = _mm_sub_pd(iz3,jz1);
 258             dx32             = _mm_sub_pd(ix3,jx2);
 259             dy32             = _mm_sub_pd(iy3,jy2);
 260             dz32             = _mm_sub_pd(iz3,jz2);
 261             dx33             = _mm_sub_pd(ix3,jx3);
 262             dy33             = _mm_sub_pd(iy3,jy3);
 263             dz33             = _mm_sub_pd(iz3,jz3);
 264
 265             /* Calculate squared distance and things based on it */
 266             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 267             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 268             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 269             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 270             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 271             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 272             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 273             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 274             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 275             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 276
 277             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 278             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 279             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 280             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 281             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 282             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 283             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 284             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 285             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 286             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 287
 288             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 289             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 290             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 291             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 292             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 293             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 294             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 295             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 296             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 297
 298             fjx0             = _mm_setzero_pd();
 299             fjy0             = _mm_setzero_pd();
 300             fjz0             = _mm_setzero_pd();
 301             fjx1             = _mm_setzero_pd();
 302             fjy1             = _mm_setzero_pd();
 303             fjz1             = _mm_setzero_pd();
 304             fjx2             = _mm_setzero_pd();
 305             fjy2             = _mm_setzero_pd();
 306             fjz2             = _mm_setzero_pd();
 307             fjx3             = _mm_setzero_pd();
 308             fjy3             = _mm_setzero_pd();
 309             fjz3             = _mm_setzero_pd();
 310
 311             /**************************
 312              * CALCULATE INTERACTIONS *
 313              **************************/
 314
 315             r00              = _mm_mul_pd(rsq00,rinv00);
 316
 317             /* Calculate table index by multiplying r with table scale and truncate to integer */
 318             rt               = _mm_mul_pd(r00,vftabscale);
 319             vfitab           = _mm_cvttpd_epi32(rt);
 320             vfeps            = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
 321             vfitab           = _mm_slli_epi32(vfitab,3);
 322
 323             /* CUBIC SPLINE TABLE DISPERSION */
 324             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 325             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
 326             GMX_MM_TRANSPOSE2_PD(Y,F);
 327             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 328             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
 329             GMX_MM_TRANSPOSE2_PD(G,H);
 330             Heps             = _mm_mul_pd(vfeps,H);
 331             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 332             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 333             vvdw6            = _mm_mul_pd(c6_00,VV);
 334             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 335             fvdw6            = _mm_mul_pd(c6_00,FF);
 336
 337             /* CUBIC SPLINE TABLE REPULSION */
 338             vfitab           = _mm_add_epi32(vfitab,ifour);
 339             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 340             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
 341             GMX_MM_TRANSPOSE2_PD(Y,F);
 342             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 343             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
 344             GMX_MM_TRANSPOSE2_PD(G,H);
 345             Heps             = _mm_mul_pd(vfeps,H);
 346             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 347             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 348             vvdw12           = _mm_mul_pd(c12_00,VV);
 349             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 350             fvdw12           = _mm_mul_pd(c12_00,FF);
 351             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 352             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 353
 354             /* Update potential sum for this i atom from the interaction with this j atom. */
 355             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 356
 357             fscal            = fvdw;
 358
 359             /* Calculate temporary vectorial force */
 360             tx               = _mm_mul_pd(fscal,dx00);
 361             ty               = _mm_mul_pd(fscal,dy00);
 362             tz               = _mm_mul_pd(fscal,dz00);
 363
 364             /* Update vectorial force */
 365             fix0             = _mm_add_pd(fix0,tx);
 366             fiy0             = _mm_add_pd(fiy0,ty);
 367             fiz0             = _mm_add_pd(fiz0,tz);
 368
 369             fjx0             = _mm_add_pd(fjx0,tx);
 370             fjy0             = _mm_add_pd(fjy0,ty);
 371             fjz0             = _mm_add_pd(fjz0,tz);
 372
 373             /**************************
 374              * CALCULATE INTERACTIONS *
 375              **************************/
 376
 377             if (gmx_mm_any_lt(rsq11,rcutoff2))
 378             {
 379
 380             /* REACTION-FIELD ELECTROSTATICS */
 381             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 382             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 383
 384             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 385
 386             /* Update potential sum for this i atom from the interaction with this j atom. */
 387             velec            = _mm_and_pd(velec,cutoff_mask);
 388             velecsum         = _mm_add_pd(velecsum,velec);
 389
 390             fscal            = felec;
 391
 392             fscal            = _mm_and_pd(fscal,cutoff_mask);
 393
 394             /* Calculate temporary vectorial force */
 395             tx               = _mm_mul_pd(fscal,dx11);
 396             ty               = _mm_mul_pd(fscal,dy11);
 397             tz               = _mm_mul_pd(fscal,dz11);
 398
 399             /* Update vectorial force */
 400             fix1             = _mm_add_pd(fix1,tx);
 401             fiy1             = _mm_add_pd(fiy1,ty);
 402             fiz1             = _mm_add_pd(fiz1,tz);
 403
 404             fjx1             = _mm_add_pd(fjx1,tx);
 405             fjy1             = _mm_add_pd(fjy1,ty);
 406             fjz1             = _mm_add_pd(fjz1,tz);
 407
 408             }
 409
 410             /**************************
 411              * CALCULATE INTERACTIONS *
 412              **************************/
 413
 414             if (gmx_mm_any_lt(rsq12,rcutoff2))
 415             {
 416
 417             /* REACTION-FIELD ELECTROSTATICS */
 418             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 419             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 420
 421             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 422
 423             /* Update potential sum for this i atom from the interaction with this j atom. */
 424             velec            = _mm_and_pd(velec,cutoff_mask);
 425             velecsum         = _mm_add_pd(velecsum,velec);
 426
 427             fscal            = felec;
 428
 429             fscal            = _mm_and_pd(fscal,cutoff_mask);
 430
 431             /* Calculate temporary vectorial force */
 432             tx               = _mm_mul_pd(fscal,dx12);
 433             ty               = _mm_mul_pd(fscal,dy12);
 434             tz               = _mm_mul_pd(fscal,dz12);
 435
 436             /* Update vectorial force */
 437             fix1             = _mm_add_pd(fix1,tx);
 438             fiy1             = _mm_add_pd(fiy1,ty);
 439             fiz1             = _mm_add_pd(fiz1,tz);
 440
 441             fjx2             = _mm_add_pd(fjx2,tx);
 442             fjy2             = _mm_add_pd(fjy2,ty);
 443             fjz2             = _mm_add_pd(fjz2,tz);
 444
 445             }
 446
 447             /**************************
 448              * CALCULATE INTERACTIONS *
 449              **************************/
 450
 451             if (gmx_mm_any_lt(rsq13,rcutoff2))
 452             {
 453
 454             /* REACTION-FIELD ELECTROSTATICS */
 455             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
 456             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
 457
 458             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 459
 460             /* Update potential sum for this i atom from the interaction with this j atom. */
 461             velec            = _mm_and_pd(velec,cutoff_mask);
 462             velecsum         = _mm_add_pd(velecsum,velec);
 463
 464             fscal            = felec;
 465
 466             fscal            = _mm_and_pd(fscal,cutoff_mask);
 467
 468             /* Calculate temporary vectorial force */
 469             tx               = _mm_mul_pd(fscal,dx13);
 470             ty               = _mm_mul_pd(fscal,dy13);
 471             tz               = _mm_mul_pd(fscal,dz13);
 472
 473             /* Update vectorial force */
 474             fix1             = _mm_add_pd(fix1,tx);
 475             fiy1             = _mm_add_pd(fiy1,ty);
 476             fiz1             = _mm_add_pd(fiz1,tz);
 477
 478             fjx3             = _mm_add_pd(fjx3,tx);
 479             fjy3             = _mm_add_pd(fjy3,ty);
 480             fjz3             = _mm_add_pd(fjz3,tz);
 481
 482             }
 483
 484             /**************************
 485              * CALCULATE INTERACTIONS *
 486              **************************/
 487
 488             if (gmx_mm_any_lt(rsq21,rcutoff2))
 489             {
 490
 491             /* REACTION-FIELD ELECTROSTATICS */
 492             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 493             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 494
 495             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 496
 497             /* Update potential sum for this i atom from the interaction with this j atom. */
 498             velec            = _mm_and_pd(velec,cutoff_mask);
 499             velecsum         = _mm_add_pd(velecsum,velec);
 500
 501             fscal            = felec;
 502
 503             fscal            = _mm_and_pd(fscal,cutoff_mask);
 504
 505             /* Calculate temporary vectorial force */
 506             tx               = _mm_mul_pd(fscal,dx21);
 507             ty               = _mm_mul_pd(fscal,dy21);
 508             tz               = _mm_mul_pd(fscal,dz21);
 509
 510             /* Update vectorial force */
 511             fix2             = _mm_add_pd(fix2,tx);
 512             fiy2             = _mm_add_pd(fiy2,ty);
 513             fiz2             = _mm_add_pd(fiz2,tz);
 514
 515             fjx1             = _mm_add_pd(fjx1,tx);
 516             fjy1             = _mm_add_pd(fjy1,ty);
 517             fjz1             = _mm_add_pd(fjz1,tz);
 518
 519             }
 520
 521             /**************************
 522              * CALCULATE INTERACTIONS *
 523              **************************/
 524
 525             if (gmx_mm_any_lt(rsq22,rcutoff2))
 526             {
 527
 528             /* REACTION-FIELD ELECTROSTATICS */
 529             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 530             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 531
 532             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 533
 534             /* Update potential sum for this i atom from the interaction with this j atom. */
 535             velec            = _mm_and_pd(velec,cutoff_mask);
 536             velecsum         = _mm_add_pd(velecsum,velec);
 537
 538             fscal            = felec;
 539
 540             fscal            = _mm_and_pd(fscal,cutoff_mask);
 541
 542             /* Calculate temporary vectorial force */
 543             tx               = _mm_mul_pd(fscal,dx22);
 544             ty               = _mm_mul_pd(fscal,dy22);
 545             tz               = _mm_mul_pd(fscal,dz22);
 546
 547             /* Update vectorial force */
 548             fix2             = _mm_add_pd(fix2,tx);
 549             fiy2             = _mm_add_pd(fiy2,ty);
 550             fiz2             = _mm_add_pd(fiz2,tz);
 551
 552             fjx2             = _mm_add_pd(fjx2,tx);
 553             fjy2             = _mm_add_pd(fjy2,ty);
 554             fjz2             = _mm_add_pd(fjz2,tz);
 555
 556             }
 557
 558             /**************************
 559              * CALCULATE INTERACTIONS *
 560              **************************/
 561
 562             if (gmx_mm_any_lt(rsq23,rcutoff2))
 563             {
 564
 565             /* REACTION-FIELD ELECTROSTATICS */
 566             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
 567             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
 568
 569             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 570
 571             /* Update potential sum for this i atom from the interaction with this j atom. */
 572             velec            = _mm_and_pd(velec,cutoff_mask);
 573             velecsum         = _mm_add_pd(velecsum,velec);
 574
 575             fscal            = felec;
 576
 577             fscal            = _mm_and_pd(fscal,cutoff_mask);
 578
 579             /* Calculate temporary vectorial force */
 580             tx               = _mm_mul_pd(fscal,dx23);
 581             ty               = _mm_mul_pd(fscal,dy23);
 582             tz               = _mm_mul_pd(fscal,dz23);
 583
 584             /* Update vectorial force */
 585             fix2             = _mm_add_pd(fix2,tx);
 586             fiy2             = _mm_add_pd(fiy2,ty);
 587             fiz2             = _mm_add_pd(fiz2,tz);
 588
 589             fjx3             = _mm_add_pd(fjx3,tx);
 590             fjy3             = _mm_add_pd(fjy3,ty);
 591             fjz3             = _mm_add_pd(fjz3,tz);
 592
 593             }
 594
 595             /**************************
 596              * CALCULATE INTERACTIONS *
 597              **************************/
 598
 599             if (gmx_mm_any_lt(rsq31,rcutoff2))
 600             {
 601
 602             /* REACTION-FIELD ELECTROSTATICS */
 603             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
 604             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
 605
 606             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
 607
 608             /* Update potential sum for this i atom from the interaction with this j atom. */
 609             velec            = _mm_and_pd(velec,cutoff_mask);
 610             velecsum         = _mm_add_pd(velecsum,velec);
 611
 612             fscal            = felec;
 613
 614             fscal            = _mm_and_pd(fscal,cutoff_mask);
 615
 616             /* Calculate temporary vectorial force */
 617             tx               = _mm_mul_pd(fscal,dx31);
 618             ty               = _mm_mul_pd(fscal,dy31);
 619             tz               = _mm_mul_pd(fscal,dz31);
 620
 621             /* Update vectorial force */
 622             fix3             = _mm_add_pd(fix3,tx);
 623             fiy3             = _mm_add_pd(fiy3,ty);
 624             fiz3             = _mm_add_pd(fiz3,tz);
 625
 626             fjx1             = _mm_add_pd(fjx1,tx);
 627             fjy1             = _mm_add_pd(fjy1,ty);
 628             fjz1             = _mm_add_pd(fjz1,tz);
 629
 630             }
 631
 632             /**************************
 633              * CALCULATE INTERACTIONS *
 634              **************************/
 635
 636             if (gmx_mm_any_lt(rsq32,rcutoff2))
 637             {
 638
 639             /* REACTION-FIELD ELECTROSTATICS */
 640             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
 641             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
 642
 643             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
 644
 645             /* Update potential sum for this i atom from the interaction with this j atom. */
 646             velec            = _mm_and_pd(velec,cutoff_mask);
 647             velecsum         = _mm_add_pd(velecsum,velec);
 648
 649             fscal            = felec;
 650
 651             fscal            = _mm_and_pd(fscal,cutoff_mask);
 652
 653             /* Calculate temporary vectorial force */
 654             tx               = _mm_mul_pd(fscal,dx32);
 655             ty               = _mm_mul_pd(fscal,dy32);
 656             tz               = _mm_mul_pd(fscal,dz32);
 657
 658             /* Update vectorial force */
 659             fix3             = _mm_add_pd(fix3,tx);
 660             fiy3             = _mm_add_pd(fiy3,ty);
 661             fiz3             = _mm_add_pd(fiz3,tz);
 662
 663             fjx2             = _mm_add_pd(fjx2,tx);
 664             fjy2             = _mm_add_pd(fjy2,ty);
 665             fjz2             = _mm_add_pd(fjz2,tz);
 666
 667             }
 668
 669             /**************************
 670              * CALCULATE INTERACTIONS *
 671              **************************/
 672
 673             if (gmx_mm_any_lt(rsq33,rcutoff2))
 674             {
 675
 676             /* REACTION-FIELD ELECTROSTATICS */
 677             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
 678             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
 679
 680             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
 681
 682             /* Update potential sum for this i atom from the interaction with this j atom. */
 683             velec            = _mm_and_pd(velec,cutoff_mask);
 684             velecsum         = _mm_add_pd(velecsum,velec);
 685
 686             fscal            = felec;
 687
 688             fscal            = _mm_and_pd(fscal,cutoff_mask);
 689
 690             /* Calculate temporary vectorial force */
 691             tx               = _mm_mul_pd(fscal,dx33);
 692             ty               = _mm_mul_pd(fscal,dy33);
 693             tz               = _mm_mul_pd(fscal,dz33);
 694
 695             /* Update vectorial force */
 696             fix3             = _mm_add_pd(fix3,tx);
 697             fiy3             = _mm_add_pd(fiy3,ty);
 698             fiz3             = _mm_add_pd(fiz3,tz);
 699
 700             fjx3             = _mm_add_pd(fjx3,tx);
 701             fjy3             = _mm_add_pd(fjy3,ty);
 702             fjz3             = _mm_add_pd(fjz3,tz);
 703
 704             }
 705
 706             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);
 707
 708             /* Inner loop uses 383 flops */
 709         }
 710
 711         if(jidx<j_index_end)
 712         {
 713
 714             jnrA             = jjnr[jidx];
 715             j_coord_offsetA  = DIM*jnrA;
 716
 717             /* load j atom coordinates */
 718             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 719                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 720                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 721
 722             /* Calculate displacement vector */
 723             dx00             = _mm_sub_pd(ix0,jx0);
 724             dy00             = _mm_sub_pd(iy0,jy0);
 725             dz00             = _mm_sub_pd(iz0,jz0);
 726             dx11             = _mm_sub_pd(ix1,jx1);
 727             dy11             = _mm_sub_pd(iy1,jy1);
 728             dz11             = _mm_sub_pd(iz1,jz1);
 729             dx12             = _mm_sub_pd(ix1,jx2);
 730             dy12             = _mm_sub_pd(iy1,jy2);
 731             dz12             = _mm_sub_pd(iz1,jz2);
 732             dx13             = _mm_sub_pd(ix1,jx3);
 733             dy13             = _mm_sub_pd(iy1,jy3);
 734             dz13             = _mm_sub_pd(iz1,jz3);
 735             dx21             = _mm_sub_pd(ix2,jx1);
 736             dy21             = _mm_sub_pd(iy2,jy1);
 737             dz21             = _mm_sub_pd(iz2,jz1);
 738             dx22             = _mm_sub_pd(ix2,jx2);
 739             dy22             = _mm_sub_pd(iy2,jy2);
 740             dz22             = _mm_sub_pd(iz2,jz2);
 741             dx23             = _mm_sub_pd(ix2,jx3);
 742             dy23             = _mm_sub_pd(iy2,jy3);
 743             dz23             = _mm_sub_pd(iz2,jz3);
 744             dx31             = _mm_sub_pd(ix3,jx1);
 745             dy31             = _mm_sub_pd(iy3,jy1);
 746             dz31             = _mm_sub_pd(iz3,jz1);
 747             dx32             = _mm_sub_pd(ix3,jx2);
 748             dy32             = _mm_sub_pd(iy3,jy2);
 749             dz32             = _mm_sub_pd(iz3,jz2);
 750             dx33             = _mm_sub_pd(ix3,jx3);
 751             dy33             = _mm_sub_pd(iy3,jy3);
 752             dz33             = _mm_sub_pd(iz3,jz3);
 753
 754             /* Calculate squared distance and things based on it */
 755             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 756             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 757             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 758             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 759             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 760             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 761             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 762             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 763             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 764             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 765
 766             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 767             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 768             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 769             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 770             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 771             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 772             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 773             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 774             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 775             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 776
 777             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 778             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 779             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 780             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 781             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 782             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 783             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 784             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 785             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 786
 787             fjx0             = _mm_setzero_pd();
 788             fjy0             = _mm_setzero_pd();
 789             fjz0             = _mm_setzero_pd();
 790             fjx1             = _mm_setzero_pd();
 791             fjy1             = _mm_setzero_pd();
 792             fjz1             = _mm_setzero_pd();
 793             fjx2             = _mm_setzero_pd();
 794             fjy2             = _mm_setzero_pd();
 795             fjz2             = _mm_setzero_pd();
 796             fjx3             = _mm_setzero_pd();
 797             fjy3             = _mm_setzero_pd();
 798             fjz3             = _mm_setzero_pd();
 799
 800             /**************************
 801              * CALCULATE INTERACTIONS *
 802              **************************/
 803
 804             r00              = _mm_mul_pd(rsq00,rinv00);
 805
 806             /* Calculate table index by multiplying r with table scale and truncate to integer */
 807             rt               = _mm_mul_pd(r00,vftabscale);
 808             vfitab           = _mm_cvttpd_epi32(rt);
 809             vfeps            = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
 810             vfitab           = _mm_slli_epi32(vfitab,3);
 811
 812             /* CUBIC SPLINE TABLE DISPERSION */
 813             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 814             F                = _mm_setzero_pd();
 815             GMX_MM_TRANSPOSE2_PD(Y,F);
 816             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 817             H                = _mm_setzero_pd();
 818             GMX_MM_TRANSPOSE2_PD(G,H);
 819             Heps             = _mm_mul_pd(vfeps,H);
 820             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 821             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 822             vvdw6            = _mm_mul_pd(c6_00,VV);
 823             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 824             fvdw6            = _mm_mul_pd(c6_00,FF);
 825
 826             /* CUBIC SPLINE TABLE REPULSION */
 827             vfitab           = _mm_add_epi32(vfitab,ifour);
 828             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 829             F                = _mm_setzero_pd();
 830             GMX_MM_TRANSPOSE2_PD(Y,F);
 831             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 832             H                = _mm_setzero_pd();
 833             GMX_MM_TRANSPOSE2_PD(G,H);
 834             Heps             = _mm_mul_pd(vfeps,H);
 835             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 836             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 837             vvdw12           = _mm_mul_pd(c12_00,VV);
 838             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 839             fvdw12           = _mm_mul_pd(c12_00,FF);
 840             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 841             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 842
 843             /* Update potential sum for this i atom from the interaction with this j atom. */
 844             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 845             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 846
 847             fscal            = fvdw;
 848
 849             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 850
 851             /* Calculate temporary vectorial force */
 852             tx               = _mm_mul_pd(fscal,dx00);
 853             ty               = _mm_mul_pd(fscal,dy00);
 854             tz               = _mm_mul_pd(fscal,dz00);
 855
 856             /* Update vectorial force */
 857             fix0             = _mm_add_pd(fix0,tx);
 858             fiy0             = _mm_add_pd(fiy0,ty);
 859             fiz0             = _mm_add_pd(fiz0,tz);
 860
 861             fjx0             = _mm_add_pd(fjx0,tx);
 862             fjy0             = _mm_add_pd(fjy0,ty);
 863             fjz0             = _mm_add_pd(fjz0,tz);
 864
 865             /**************************
 866              * CALCULATE INTERACTIONS *
 867              **************************/
 868
 869             if (gmx_mm_any_lt(rsq11,rcutoff2))
 870             {
 871
 872             /* REACTION-FIELD ELECTROSTATICS */
 873             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 874             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 875
 876             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 877
 878             /* Update potential sum for this i atom from the interaction with this j atom. */
 879             velec            = _mm_and_pd(velec,cutoff_mask);
 880             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 881             velecsum         = _mm_add_pd(velecsum,velec);
 882
 883             fscal            = felec;
 884
 885             fscal            = _mm_and_pd(fscal,cutoff_mask);
 886
 887             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 888
 889             /* Calculate temporary vectorial force */
 890             tx               = _mm_mul_pd(fscal,dx11);
 891             ty               = _mm_mul_pd(fscal,dy11);
 892             tz               = _mm_mul_pd(fscal,dz11);
 893
 894             /* Update vectorial force */
 895             fix1             = _mm_add_pd(fix1,tx);
 896             fiy1             = _mm_add_pd(fiy1,ty);
 897             fiz1             = _mm_add_pd(fiz1,tz);
 898
 899             fjx1             = _mm_add_pd(fjx1,tx);
 900             fjy1             = _mm_add_pd(fjy1,ty);
 901             fjz1             = _mm_add_pd(fjz1,tz);
 902
 903             }
 904
 905             /**************************
 906              * CALCULATE INTERACTIONS *
 907              **************************/
 908
 909             if (gmx_mm_any_lt(rsq12,rcutoff2))
 910             {
 911
 912             /* REACTION-FIELD ELECTROSTATICS */
 913             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 914             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 915
 916             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 917
 918             /* Update potential sum for this i atom from the interaction with this j atom. */
 919             velec            = _mm_and_pd(velec,cutoff_mask);
 920             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 921             velecsum         = _mm_add_pd(velecsum,velec);
 922
 923             fscal            = felec;
 924
 925             fscal            = _mm_and_pd(fscal,cutoff_mask);
 926
 927             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 928
 929             /* Calculate temporary vectorial force */
 930             tx               = _mm_mul_pd(fscal,dx12);
 931             ty               = _mm_mul_pd(fscal,dy12);
 932             tz               = _mm_mul_pd(fscal,dz12);
 933
 934             /* Update vectorial force */
 935             fix1             = _mm_add_pd(fix1,tx);
 936             fiy1             = _mm_add_pd(fiy1,ty);
 937             fiz1             = _mm_add_pd(fiz1,tz);
 938
 939             fjx2             = _mm_add_pd(fjx2,tx);
 940             fjy2             = _mm_add_pd(fjy2,ty);
 941             fjz2             = _mm_add_pd(fjz2,tz);
 942
 943             }
 944
 945             /**************************
 946              * CALCULATE INTERACTIONS *
 947              **************************/
 948
 949             if (gmx_mm_any_lt(rsq13,rcutoff2))
 950             {
 951
 952             /* REACTION-FIELD ELECTROSTATICS */
 953             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_add_pd(rinv13,_mm_mul_pd(krf,rsq13)),crf));
 954             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
 955
 956             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 957
 958             /* Update potential sum for this i atom from the interaction with this j atom. */
 959             velec            = _mm_and_pd(velec,cutoff_mask);
 960             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 961             velecsum         = _mm_add_pd(velecsum,velec);
 962
 963             fscal            = felec;
 964
 965             fscal            = _mm_and_pd(fscal,cutoff_mask);
 966
 967             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 968
 969             /* Calculate temporary vectorial force */
 970             tx               = _mm_mul_pd(fscal,dx13);
 971             ty               = _mm_mul_pd(fscal,dy13);
 972             tz               = _mm_mul_pd(fscal,dz13);
 973
 974             /* Update vectorial force */
 975             fix1             = _mm_add_pd(fix1,tx);
 976             fiy1             = _mm_add_pd(fiy1,ty);
 977             fiz1             = _mm_add_pd(fiz1,tz);
 978
 979             fjx3             = _mm_add_pd(fjx3,tx);
 980             fjy3             = _mm_add_pd(fjy3,ty);
 981             fjz3             = _mm_add_pd(fjz3,tz);
 982
 983             }
 984
 985             /**************************
 986              * CALCULATE INTERACTIONS *
 987              **************************/
 988
 989             if (gmx_mm_any_lt(rsq21,rcutoff2))
 990             {
 991
 992             /* REACTION-FIELD ELECTROSTATICS */
 993             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 994             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 995
 996             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 997
 998             /* Update potential sum for this i atom from the interaction with this j atom. */
 999             velec            = _mm_and_pd(velec,cutoff_mask);
1000             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1001             velecsum         = _mm_add_pd(velecsum,velec);
1002
1003             fscal            = felec;
1004
1005             fscal            = _mm_and_pd(fscal,cutoff_mask);
1006
1007             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1008
1009             /* Calculate temporary vectorial force */
1010             tx               = _mm_mul_pd(fscal,dx21);
1011             ty               = _mm_mul_pd(fscal,dy21);
1012             tz               = _mm_mul_pd(fscal,dz21);
1013
1014             /* Update vectorial force */
1015             fix2             = _mm_add_pd(fix2,tx);
1016             fiy2             = _mm_add_pd(fiy2,ty);
1017             fiz2             = _mm_add_pd(fiz2,tz);
1018
1019             fjx1             = _mm_add_pd(fjx1,tx);
1020             fjy1             = _mm_add_pd(fjy1,ty);
1021             fjz1             = _mm_add_pd(fjz1,tz);
1022
1023             }
1024
1025             /**************************
1026              * CALCULATE INTERACTIONS *
1027              **************************/
1028
1029             if (gmx_mm_any_lt(rsq22,rcutoff2))
1030             {
1031
1032             /* REACTION-FIELD ELECTROSTATICS */
1033             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
1034             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1035
1036             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1037
1038             /* Update potential sum for this i atom from the interaction with this j atom. */
1039             velec            = _mm_and_pd(velec,cutoff_mask);
1040             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1041             velecsum         = _mm_add_pd(velecsum,velec);
1042
1043             fscal            = felec;
1044
1045             fscal            = _mm_and_pd(fscal,cutoff_mask);
1046
1047             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1048
1049             /* Calculate temporary vectorial force */
1050             tx               = _mm_mul_pd(fscal,dx22);
1051             ty               = _mm_mul_pd(fscal,dy22);
1052             tz               = _mm_mul_pd(fscal,dz22);
1053
1054             /* Update vectorial force */
1055             fix2             = _mm_add_pd(fix2,tx);
1056             fiy2             = _mm_add_pd(fiy2,ty);
1057             fiz2             = _mm_add_pd(fiz2,tz);
1058
1059             fjx2             = _mm_add_pd(fjx2,tx);
1060             fjy2             = _mm_add_pd(fjy2,ty);
1061             fjz2             = _mm_add_pd(fjz2,tz);
1062
1063             }
1064
1065             /**************************
1066              * CALCULATE INTERACTIONS *
1067              **************************/
1068
1069             if (gmx_mm_any_lt(rsq23,rcutoff2))
1070             {
1071
1072             /* REACTION-FIELD ELECTROSTATICS */
1073             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_add_pd(rinv23,_mm_mul_pd(krf,rsq23)),crf));
1074             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1075
1076             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1077
1078             /* Update potential sum for this i atom from the interaction with this j atom. */
1079             velec            = _mm_and_pd(velec,cutoff_mask);
1080             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1081             velecsum         = _mm_add_pd(velecsum,velec);
1082
1083             fscal            = felec;
1084
1085             fscal            = _mm_and_pd(fscal,cutoff_mask);
1086
1087             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1088
1089             /* Calculate temporary vectorial force */
1090             tx               = _mm_mul_pd(fscal,dx23);
1091             ty               = _mm_mul_pd(fscal,dy23);
1092             tz               = _mm_mul_pd(fscal,dz23);
1093
1094             /* Update vectorial force */
1095             fix2             = _mm_add_pd(fix2,tx);
1096             fiy2             = _mm_add_pd(fiy2,ty);
1097             fiz2             = _mm_add_pd(fiz2,tz);
1098
1099             fjx3             = _mm_add_pd(fjx3,tx);
1100             fjy3             = _mm_add_pd(fjy3,ty);
1101             fjz3             = _mm_add_pd(fjz3,tz);
1102
1103             }
1104
1105             /**************************
1106              * CALCULATE INTERACTIONS *
1107              **************************/
1108
1109             if (gmx_mm_any_lt(rsq31,rcutoff2))
1110             {
1111
1112             /* REACTION-FIELD ELECTROSTATICS */
1113             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_add_pd(rinv31,_mm_mul_pd(krf,rsq31)),crf));
1114             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1115
1116             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1117
1118             /* Update potential sum for this i atom from the interaction with this j atom. */
1119             velec            = _mm_and_pd(velec,cutoff_mask);
1120             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1121             velecsum         = _mm_add_pd(velecsum,velec);
1122
1123             fscal            = felec;
1124
1125             fscal            = _mm_and_pd(fscal,cutoff_mask);
1126
1127             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1128
1129             /* Calculate temporary vectorial force */
1130             tx               = _mm_mul_pd(fscal,dx31);
1131             ty               = _mm_mul_pd(fscal,dy31);
1132             tz               = _mm_mul_pd(fscal,dz31);
1133
1134             /* Update vectorial force */
1135             fix3             = _mm_add_pd(fix3,tx);
1136             fiy3             = _mm_add_pd(fiy3,ty);
1137             fiz3             = _mm_add_pd(fiz3,tz);
1138
1139             fjx1             = _mm_add_pd(fjx1,tx);
1140             fjy1             = _mm_add_pd(fjy1,ty);
1141             fjz1             = _mm_add_pd(fjz1,tz);
1142
1143             }
1144
1145             /**************************
1146              * CALCULATE INTERACTIONS *
1147              **************************/
1148
1149             if (gmx_mm_any_lt(rsq32,rcutoff2))
1150             {
1151
1152             /* REACTION-FIELD ELECTROSTATICS */
1153             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_add_pd(rinv32,_mm_mul_pd(krf,rsq32)),crf));
1154             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1155
1156             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1157
1158             /* Update potential sum for this i atom from the interaction with this j atom. */
1159             velec            = _mm_and_pd(velec,cutoff_mask);
1160             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1161             velecsum         = _mm_add_pd(velecsum,velec);
1162
1163             fscal            = felec;
1164
1165             fscal            = _mm_and_pd(fscal,cutoff_mask);
1166
1167             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1168
1169             /* Calculate temporary vectorial force */
1170             tx               = _mm_mul_pd(fscal,dx32);
1171             ty               = _mm_mul_pd(fscal,dy32);
1172             tz               = _mm_mul_pd(fscal,dz32);
1173
1174             /* Update vectorial force */
1175             fix3             = _mm_add_pd(fix3,tx);
1176             fiy3             = _mm_add_pd(fiy3,ty);
1177             fiz3             = _mm_add_pd(fiz3,tz);
1178
1179             fjx2             = _mm_add_pd(fjx2,tx);
1180             fjy2             = _mm_add_pd(fjy2,ty);
1181             fjz2             = _mm_add_pd(fjz2,tz);
1182
1183             }
1184
1185             /**************************
1186              * CALCULATE INTERACTIONS *
1187              **************************/
1188
1189             if (gmx_mm_any_lt(rsq33,rcutoff2))
1190             {
1191
1192             /* REACTION-FIELD ELECTROSTATICS */
1193             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_add_pd(rinv33,_mm_mul_pd(krf,rsq33)),crf));
1194             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1195
1196             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1197
1198             /* Update potential sum for this i atom from the interaction with this j atom. */
1199             velec            = _mm_and_pd(velec,cutoff_mask);
1200             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1201             velecsum         = _mm_add_pd(velecsum,velec);
1202
1203             fscal            = felec;
1204
1205             fscal            = _mm_and_pd(fscal,cutoff_mask);
1206
1207             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1208
1209             /* Calculate temporary vectorial force */
1210             tx               = _mm_mul_pd(fscal,dx33);
1211             ty               = _mm_mul_pd(fscal,dy33);
1212             tz               = _mm_mul_pd(fscal,dz33);
1213
1214             /* Update vectorial force */
1215             fix3             = _mm_add_pd(fix3,tx);
1216             fiy3             = _mm_add_pd(fiy3,ty);
1217             fiz3             = _mm_add_pd(fiz3,tz);
1218
1219             fjx3             = _mm_add_pd(fjx3,tx);
1220             fjy3             = _mm_add_pd(fjy3,ty);
1221             fjz3             = _mm_add_pd(fjz3,tz);
1222
1223             }
1224
1225             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1226
1227             /* Inner loop uses 383 flops */
1228         }
1229
1230         /* End of innermost loop */
1231
1232         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1233                                               f+i_coord_offset,fshift+i_shift_offset);
1234
1235         ggid                        = gid[iidx];
1236         /* Update potential energies */
1237         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1238         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
1239
1240         /* Increment number of inner iterations */
1241         inneriter                  += j_index_end - j_index_start;
1242
1243         /* Outer loop uses 26 flops */
1244     }
1245
1246     /* Increment number of outer iterations */
1247     outeriter        += nri;
1248
1249     /* Update outer/inner flops */
1250
1251     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*383);
1252 }
1253 /*
1254  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW4W4_F_sse2_double
1255  * Electrostatics interaction: ReactionField
1256  * VdW interaction:            CubicSplineTable
1257  * Geometry:                   Water4-Water4
1258  * Calculate force/pot:        Force
1259  */
1260 void
1261 nb_kernel_ElecRFCut_VdwCSTab_GeomW4W4_F_sse2_double
1262                     (t_nblist                    * gmx_restrict       nlist,
1263                      rvec                        * gmx_restrict          xx,
1264                      rvec                        * gmx_restrict          ff,
1265                      t_forcerec                  * gmx_restrict          fr,
1266                      t_mdatoms                   * gmx_restrict     mdatoms,
1267                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1268                      t_nrnb                      * gmx_restrict        nrnb)
1269 {
1270     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1271      * just 0 for non-waters.
1272      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1273      * jnr indices corresponding to data put in the four positions in the SIMD register.
1274      */
1275     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1276     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1277     int              jnrA,jnrB;
1278     int              j_coord_offsetA,j_coord_offsetB;
1279     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1280     real             rcutoff_scalar;
1281     real             *shiftvec,*fshift,*x,*f;
1282     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1283     int              vdwioffset0;
1284     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1285     int              vdwioffset1;
1286     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1287     int              vdwioffset2;
1288     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1289     int              vdwioffset3;
1290     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1291     int              vdwjidx0A,vdwjidx0B;
1292     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1293     int              vdwjidx1A,vdwjidx1B;
1294     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1295     int              vdwjidx2A,vdwjidx2B;
1296     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1297     int              vdwjidx3A,vdwjidx3B;
1298     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1299     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1300     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1301     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1302     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1303     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1304     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1305     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1306     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1307     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1308     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1309     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1310     real             *charge;
1311     int              nvdwtype;
1312     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1313     int              *vdwtype;
1314     real             *vdwparam;
1315     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1316     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1317     __m128i          vfitab;
1318     __m128i          ifour       = _mm_set1_epi32(4);
1319     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
1320     real             *vftab;
1321     __m128d          dummy_mask,cutoff_mask;
1322     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1323     __m128d          one     = _mm_set1_pd(1.0);
1324     __m128d          two     = _mm_set1_pd(2.0);
1325     x                = xx[0];
1326     f                = ff[0];
1327
1328     nri              = nlist->nri;
1329     iinr             = nlist->iinr;
1330     jindex           = nlist->jindex;
1331     jjnr             = nlist->jjnr;
1332     shiftidx         = nlist->shift;
1333     gid              = nlist->gid;
1334     shiftvec         = fr->shift_vec[0];
1335     fshift           = fr->fshift[0];
1336     facel            = _mm_set1_pd(fr->epsfac);
1337     charge           = mdatoms->chargeA;
1338     krf              = _mm_set1_pd(fr->ic->k_rf);
1339     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1340     crf              = _mm_set1_pd(fr->ic->c_rf);
1341     nvdwtype         = fr->ntype;
1342     vdwparam         = fr->nbfp;
1343     vdwtype          = mdatoms->typeA;
1344
1345     vftab            = kernel_data->table_vdw->data;
1346     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
1347
1348     /* Setup water-specific parameters */
1349     inr              = nlist->iinr[0];
1350     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1351     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1352     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1353     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1354
1355     jq1              = _mm_set1_pd(charge[inr+1]);
1356     jq2              = _mm_set1_pd(charge[inr+2]);
1357     jq3              = _mm_set1_pd(charge[inr+3]);
1358     vdwjidx0A        = 2*vdwtype[inr+0];
1359     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1360     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1361     qq11             = _mm_mul_pd(iq1,jq1);
1362     qq12             = _mm_mul_pd(iq1,jq2);
1363     qq13             = _mm_mul_pd(iq1,jq3);
1364     qq21             = _mm_mul_pd(iq2,jq1);
1365     qq22             = _mm_mul_pd(iq2,jq2);
1366     qq23             = _mm_mul_pd(iq2,jq3);
1367     qq31             = _mm_mul_pd(iq3,jq1);
1368     qq32             = _mm_mul_pd(iq3,jq2);
1369     qq33             = _mm_mul_pd(iq3,jq3);
1370
1371     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1372     rcutoff_scalar   = fr->rcoulomb;
1373     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1374     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1375
1376     /* Avoid stupid compiler warnings */
1377     jnrA = jnrB = 0;
1378     j_coord_offsetA = 0;
1379     j_coord_offsetB = 0;
1380
1381     outeriter        = 0;
1382     inneriter        = 0;
1383
1384     /* Start outer loop over neighborlists */
1385     for(iidx=0; iidx<nri; iidx++)
1386     {
1387         /* Load shift vector for this list */
1388         i_shift_offset   = DIM*shiftidx[iidx];
1389
1390         /* Load limits for loop over neighbors */
1391         j_index_start    = jindex[iidx];
1392         j_index_end      = jindex[iidx+1];
1393
1394         /* Get outer coordinate index */
1395         inr              = iinr[iidx];
1396         i_coord_offset   = DIM*inr;
1397
1398         /* Load i particle coords and add shift vector */
1399         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1400                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1401
1402         fix0             = _mm_setzero_pd();
1403         fiy0             = _mm_setzero_pd();
1404         fiz0             = _mm_setzero_pd();
1405         fix1             = _mm_setzero_pd();
1406         fiy1             = _mm_setzero_pd();
1407         fiz1             = _mm_setzero_pd();
1408         fix2             = _mm_setzero_pd();
1409         fiy2             = _mm_setzero_pd();
1410         fiz2             = _mm_setzero_pd();
1411         fix3             = _mm_setzero_pd();
1412         fiy3             = _mm_setzero_pd();
1413         fiz3             = _mm_setzero_pd();
1414
1415         /* Start inner kernel loop */
1416         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1417         {
1418
1419             /* Get j neighbor index, and coordinate index */
1420             jnrA             = jjnr[jidx];
1421             jnrB             = jjnr[jidx+1];
1422             j_coord_offsetA  = DIM*jnrA;
1423             j_coord_offsetB  = DIM*jnrB;
1424
1425             /* load j atom coordinates */
1426             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1427                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1428                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1429
1430             /* Calculate displacement vector */
1431             dx00             = _mm_sub_pd(ix0,jx0);
1432             dy00             = _mm_sub_pd(iy0,jy0);
1433             dz00             = _mm_sub_pd(iz0,jz0);
1434             dx11             = _mm_sub_pd(ix1,jx1);
1435             dy11             = _mm_sub_pd(iy1,jy1);
1436             dz11             = _mm_sub_pd(iz1,jz1);
1437             dx12             = _mm_sub_pd(ix1,jx2);
1438             dy12             = _mm_sub_pd(iy1,jy2);
1439             dz12             = _mm_sub_pd(iz1,jz2);
1440             dx13             = _mm_sub_pd(ix1,jx3);
1441             dy13             = _mm_sub_pd(iy1,jy3);
1442             dz13             = _mm_sub_pd(iz1,jz3);
1443             dx21             = _mm_sub_pd(ix2,jx1);
1444             dy21             = _mm_sub_pd(iy2,jy1);
1445             dz21             = _mm_sub_pd(iz2,jz1);
1446             dx22             = _mm_sub_pd(ix2,jx2);
1447             dy22             = _mm_sub_pd(iy2,jy2);
1448             dz22             = _mm_sub_pd(iz2,jz2);
1449             dx23             = _mm_sub_pd(ix2,jx3);
1450             dy23             = _mm_sub_pd(iy2,jy3);
1451             dz23             = _mm_sub_pd(iz2,jz3);
1452             dx31             = _mm_sub_pd(ix3,jx1);
1453             dy31             = _mm_sub_pd(iy3,jy1);
1454             dz31             = _mm_sub_pd(iz3,jz1);
1455             dx32             = _mm_sub_pd(ix3,jx2);
1456             dy32             = _mm_sub_pd(iy3,jy2);
1457             dz32             = _mm_sub_pd(iz3,jz2);
1458             dx33             = _mm_sub_pd(ix3,jx3);
1459             dy33             = _mm_sub_pd(iy3,jy3);
1460             dz33             = _mm_sub_pd(iz3,jz3);
1461
1462             /* Calculate squared distance and things based on it */
1463             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1464             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1465             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1466             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1467             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1468             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1469             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1470             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1471             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1472             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1473
1474             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1475             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1476             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1477             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1478             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1479             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1480             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1481             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1482             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1483             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1484
1485             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1486             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1487             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1488             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1489             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1490             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1491             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1492             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1493             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1494
1495             fjx0             = _mm_setzero_pd();
1496             fjy0             = _mm_setzero_pd();
1497             fjz0             = _mm_setzero_pd();
1498             fjx1             = _mm_setzero_pd();
1499             fjy1             = _mm_setzero_pd();
1500             fjz1             = _mm_setzero_pd();
1501             fjx2             = _mm_setzero_pd();
1502             fjy2             = _mm_setzero_pd();
1503             fjz2             = _mm_setzero_pd();
1504             fjx3             = _mm_setzero_pd();
1505             fjy3             = _mm_setzero_pd();
1506             fjz3             = _mm_setzero_pd();
1507
1508             /**************************
1509              * CALCULATE INTERACTIONS *
1510              **************************/
1511
1512             r00              = _mm_mul_pd(rsq00,rinv00);
1513
1514             /* Calculate table index by multiplying r with table scale and truncate to integer */
1515             rt               = _mm_mul_pd(r00,vftabscale);
1516             vfitab           = _mm_cvttpd_epi32(rt);
1517             vfeps            = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
1518             vfitab           = _mm_slli_epi32(vfitab,3);
1519
1520             /* CUBIC SPLINE TABLE DISPERSION */
1521             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1522             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
1523             GMX_MM_TRANSPOSE2_PD(Y,F);
1524             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1525             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
1526             GMX_MM_TRANSPOSE2_PD(G,H);
1527             Heps             = _mm_mul_pd(vfeps,H);
1528             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1529             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1530             fvdw6            = _mm_mul_pd(c6_00,FF);
1531
1532             /* CUBIC SPLINE TABLE REPULSION */
1533             vfitab           = _mm_add_epi32(vfitab,ifour);
1534             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1535             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
1536             GMX_MM_TRANSPOSE2_PD(Y,F);
1537             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1538             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
1539             GMX_MM_TRANSPOSE2_PD(G,H);
1540             Heps             = _mm_mul_pd(vfeps,H);
1541             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1542             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1543             fvdw12           = _mm_mul_pd(c12_00,FF);
1544             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1545
1546             fscal            = fvdw;
1547
1548             /* Calculate temporary vectorial force */
1549             tx               = _mm_mul_pd(fscal,dx00);
1550             ty               = _mm_mul_pd(fscal,dy00);
1551             tz               = _mm_mul_pd(fscal,dz00);
1552
1553             /* Update vectorial force */
1554             fix0             = _mm_add_pd(fix0,tx);
1555             fiy0             = _mm_add_pd(fiy0,ty);
1556             fiz0             = _mm_add_pd(fiz0,tz);
1557
1558             fjx0             = _mm_add_pd(fjx0,tx);
1559             fjy0             = _mm_add_pd(fjy0,ty);
1560             fjz0             = _mm_add_pd(fjz0,tz);
1561
1562             /**************************
1563              * CALCULATE INTERACTIONS *
1564              **************************/
1565
1566             if (gmx_mm_any_lt(rsq11,rcutoff2))
1567             {
1568
1569             /* REACTION-FIELD ELECTROSTATICS */
1570             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1571
1572             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1573
1574             fscal            = felec;
1575
1576             fscal            = _mm_and_pd(fscal,cutoff_mask);
1577
1578             /* Calculate temporary vectorial force */
1579             tx               = _mm_mul_pd(fscal,dx11);
1580             ty               = _mm_mul_pd(fscal,dy11);
1581             tz               = _mm_mul_pd(fscal,dz11);
1582
1583             /* Update vectorial force */
1584             fix1             = _mm_add_pd(fix1,tx);
1585             fiy1             = _mm_add_pd(fiy1,ty);
1586             fiz1             = _mm_add_pd(fiz1,tz);
1587
1588             fjx1             = _mm_add_pd(fjx1,tx);
1589             fjy1             = _mm_add_pd(fjy1,ty);
1590             fjz1             = _mm_add_pd(fjz1,tz);
1591
1592             }
1593
1594             /**************************
1595              * CALCULATE INTERACTIONS *
1596              **************************/
1597
1598             if (gmx_mm_any_lt(rsq12,rcutoff2))
1599             {
1600
1601             /* REACTION-FIELD ELECTROSTATICS */
1602             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1603
1604             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1605
1606             fscal            = felec;
1607
1608             fscal            = _mm_and_pd(fscal,cutoff_mask);
1609
1610             /* Calculate temporary vectorial force */
1611             tx               = _mm_mul_pd(fscal,dx12);
1612             ty               = _mm_mul_pd(fscal,dy12);
1613             tz               = _mm_mul_pd(fscal,dz12);
1614
1615             /* Update vectorial force */
1616             fix1             = _mm_add_pd(fix1,tx);
1617             fiy1             = _mm_add_pd(fiy1,ty);
1618             fiz1             = _mm_add_pd(fiz1,tz);
1619
1620             fjx2             = _mm_add_pd(fjx2,tx);
1621             fjy2             = _mm_add_pd(fjy2,ty);
1622             fjz2             = _mm_add_pd(fjz2,tz);
1623
1624             }
1625
1626             /**************************
1627              * CALCULATE INTERACTIONS *
1628              **************************/
1629
1630             if (gmx_mm_any_lt(rsq13,rcutoff2))
1631             {
1632
1633             /* REACTION-FIELD ELECTROSTATICS */
1634             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
1635
1636             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1637
1638             fscal            = felec;
1639
1640             fscal            = _mm_and_pd(fscal,cutoff_mask);
1641
1642             /* Calculate temporary vectorial force */
1643             tx               = _mm_mul_pd(fscal,dx13);
1644             ty               = _mm_mul_pd(fscal,dy13);
1645             tz               = _mm_mul_pd(fscal,dz13);
1646
1647             /* Update vectorial force */
1648             fix1             = _mm_add_pd(fix1,tx);
1649             fiy1             = _mm_add_pd(fiy1,ty);
1650             fiz1             = _mm_add_pd(fiz1,tz);
1651
1652             fjx3             = _mm_add_pd(fjx3,tx);
1653             fjy3             = _mm_add_pd(fjy3,ty);
1654             fjz3             = _mm_add_pd(fjz3,tz);
1655
1656             }
1657
1658             /**************************
1659              * CALCULATE INTERACTIONS *
1660              **************************/
1661
1662             if (gmx_mm_any_lt(rsq21,rcutoff2))
1663             {
1664
1665             /* REACTION-FIELD ELECTROSTATICS */
1666             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1667
1668             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1669
1670             fscal            = felec;
1671
1672             fscal            = _mm_and_pd(fscal,cutoff_mask);
1673
1674             /* Calculate temporary vectorial force */
1675             tx               = _mm_mul_pd(fscal,dx21);
1676             ty               = _mm_mul_pd(fscal,dy21);
1677             tz               = _mm_mul_pd(fscal,dz21);
1678
1679             /* Update vectorial force */
1680             fix2             = _mm_add_pd(fix2,tx);
1681             fiy2             = _mm_add_pd(fiy2,ty);
1682             fiz2             = _mm_add_pd(fiz2,tz);
1683
1684             fjx1             = _mm_add_pd(fjx1,tx);
1685             fjy1             = _mm_add_pd(fjy1,ty);
1686             fjz1             = _mm_add_pd(fjz1,tz);
1687
1688             }
1689
1690             /**************************
1691              * CALCULATE INTERACTIONS *
1692              **************************/
1693
1694             if (gmx_mm_any_lt(rsq22,rcutoff2))
1695             {
1696
1697             /* REACTION-FIELD ELECTROSTATICS */
1698             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1699
1700             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1701
1702             fscal            = felec;
1703
1704             fscal            = _mm_and_pd(fscal,cutoff_mask);
1705
1706             /* Calculate temporary vectorial force */
1707             tx               = _mm_mul_pd(fscal,dx22);
1708             ty               = _mm_mul_pd(fscal,dy22);
1709             tz               = _mm_mul_pd(fscal,dz22);
1710
1711             /* Update vectorial force */
1712             fix2             = _mm_add_pd(fix2,tx);
1713             fiy2             = _mm_add_pd(fiy2,ty);
1714             fiz2             = _mm_add_pd(fiz2,tz);
1715
1716             fjx2             = _mm_add_pd(fjx2,tx);
1717             fjy2             = _mm_add_pd(fjy2,ty);
1718             fjz2             = _mm_add_pd(fjz2,tz);
1719
1720             }
1721
1722             /**************************
1723              * CALCULATE INTERACTIONS *
1724              **************************/
1725
1726             if (gmx_mm_any_lt(rsq23,rcutoff2))
1727             {
1728
1729             /* REACTION-FIELD ELECTROSTATICS */
1730             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
1731
1732             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1733
1734             fscal            = felec;
1735
1736             fscal            = _mm_and_pd(fscal,cutoff_mask);
1737
1738             /* Calculate temporary vectorial force */
1739             tx               = _mm_mul_pd(fscal,dx23);
1740             ty               = _mm_mul_pd(fscal,dy23);
1741             tz               = _mm_mul_pd(fscal,dz23);
1742
1743             /* Update vectorial force */
1744             fix2             = _mm_add_pd(fix2,tx);
1745             fiy2             = _mm_add_pd(fiy2,ty);
1746             fiz2             = _mm_add_pd(fiz2,tz);
1747
1748             fjx3             = _mm_add_pd(fjx3,tx);
1749             fjy3             = _mm_add_pd(fjy3,ty);
1750             fjz3             = _mm_add_pd(fjz3,tz);
1751
1752             }
1753
1754             /**************************
1755              * CALCULATE INTERACTIONS *
1756              **************************/
1757
1758             if (gmx_mm_any_lt(rsq31,rcutoff2))
1759             {
1760
1761             /* REACTION-FIELD ELECTROSTATICS */
1762             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
1763
1764             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1765
1766             fscal            = felec;
1767
1768             fscal            = _mm_and_pd(fscal,cutoff_mask);
1769
1770             /* Calculate temporary vectorial force */
1771             tx               = _mm_mul_pd(fscal,dx31);
1772             ty               = _mm_mul_pd(fscal,dy31);
1773             tz               = _mm_mul_pd(fscal,dz31);
1774
1775             /* Update vectorial force */
1776             fix3             = _mm_add_pd(fix3,tx);
1777             fiy3             = _mm_add_pd(fiy3,ty);
1778             fiz3             = _mm_add_pd(fiz3,tz);
1779
1780             fjx1             = _mm_add_pd(fjx1,tx);
1781             fjy1             = _mm_add_pd(fjy1,ty);
1782             fjz1             = _mm_add_pd(fjz1,tz);
1783
1784             }
1785
1786             /**************************
1787              * CALCULATE INTERACTIONS *
1788              **************************/
1789
1790             if (gmx_mm_any_lt(rsq32,rcutoff2))
1791             {
1792
1793             /* REACTION-FIELD ELECTROSTATICS */
1794             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
1795
1796             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1797
1798             fscal            = felec;
1799
1800             fscal            = _mm_and_pd(fscal,cutoff_mask);
1801
1802             /* Calculate temporary vectorial force */
1803             tx               = _mm_mul_pd(fscal,dx32);
1804             ty               = _mm_mul_pd(fscal,dy32);
1805             tz               = _mm_mul_pd(fscal,dz32);
1806
1807             /* Update vectorial force */
1808             fix3             = _mm_add_pd(fix3,tx);
1809             fiy3             = _mm_add_pd(fiy3,ty);
1810             fiz3             = _mm_add_pd(fiz3,tz);
1811
1812             fjx2             = _mm_add_pd(fjx2,tx);
1813             fjy2             = _mm_add_pd(fjy2,ty);
1814             fjz2             = _mm_add_pd(fjz2,tz);
1815
1816             }
1817
1818             /**************************
1819              * CALCULATE INTERACTIONS *
1820              **************************/
1821
1822             if (gmx_mm_any_lt(rsq33,rcutoff2))
1823             {
1824
1825             /* REACTION-FIELD ELECTROSTATICS */
1826             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
1827
1828             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1829
1830             fscal            = felec;
1831
1832             fscal            = _mm_and_pd(fscal,cutoff_mask);
1833
1834             /* Calculate temporary vectorial force */
1835             tx               = _mm_mul_pd(fscal,dx33);
1836             ty               = _mm_mul_pd(fscal,dy33);
1837             tz               = _mm_mul_pd(fscal,dz33);
1838
1839             /* Update vectorial force */
1840             fix3             = _mm_add_pd(fix3,tx);
1841             fiy3             = _mm_add_pd(fiy3,ty);
1842             fiz3             = _mm_add_pd(fiz3,tz);
1843
1844             fjx3             = _mm_add_pd(fjx3,tx);
1845             fjy3             = _mm_add_pd(fjy3,ty);
1846             fjz3             = _mm_add_pd(fjz3,tz);
1847
1848             }
1849
1850             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);
1851
1852             /* Inner loop uses 321 flops */
1853         }
1854
1855         if(jidx<j_index_end)
1856         {
1857
1858             jnrA             = jjnr[jidx];
1859             j_coord_offsetA  = DIM*jnrA;
1860
1861             /* load j atom coordinates */
1862             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1863                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1864                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1865
1866             /* Calculate displacement vector */
1867             dx00             = _mm_sub_pd(ix0,jx0);
1868             dy00             = _mm_sub_pd(iy0,jy0);
1869             dz00             = _mm_sub_pd(iz0,jz0);
1870             dx11             = _mm_sub_pd(ix1,jx1);
1871             dy11             = _mm_sub_pd(iy1,jy1);
1872             dz11             = _mm_sub_pd(iz1,jz1);
1873             dx12             = _mm_sub_pd(ix1,jx2);
1874             dy12             = _mm_sub_pd(iy1,jy2);
1875             dz12             = _mm_sub_pd(iz1,jz2);
1876             dx13             = _mm_sub_pd(ix1,jx3);
1877             dy13             = _mm_sub_pd(iy1,jy3);
1878             dz13             = _mm_sub_pd(iz1,jz3);
1879             dx21             = _mm_sub_pd(ix2,jx1);
1880             dy21             = _mm_sub_pd(iy2,jy1);
1881             dz21             = _mm_sub_pd(iz2,jz1);
1882             dx22             = _mm_sub_pd(ix2,jx2);
1883             dy22             = _mm_sub_pd(iy2,jy2);
1884             dz22             = _mm_sub_pd(iz2,jz2);
1885             dx23             = _mm_sub_pd(ix2,jx3);
1886             dy23             = _mm_sub_pd(iy2,jy3);
1887             dz23             = _mm_sub_pd(iz2,jz3);
1888             dx31             = _mm_sub_pd(ix3,jx1);
1889             dy31             = _mm_sub_pd(iy3,jy1);
1890             dz31             = _mm_sub_pd(iz3,jz1);
1891             dx32             = _mm_sub_pd(ix3,jx2);
1892             dy32             = _mm_sub_pd(iy3,jy2);
1893             dz32             = _mm_sub_pd(iz3,jz2);
1894             dx33             = _mm_sub_pd(ix3,jx3);
1895             dy33             = _mm_sub_pd(iy3,jy3);
1896             dz33             = _mm_sub_pd(iz3,jz3);
1897
1898             /* Calculate squared distance and things based on it */
1899             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1900             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1901             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1902             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1903             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1904             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1905             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1906             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1907             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1908             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1909
1910             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1911             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1912             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1913             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1914             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1915             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1916             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1917             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1918             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1919             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1920
1921             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1922             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1923             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1924             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1925             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1926             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1927             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1928             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1929             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1930
1931             fjx0             = _mm_setzero_pd();
1932             fjy0             = _mm_setzero_pd();
1933             fjz0             = _mm_setzero_pd();
1934             fjx1             = _mm_setzero_pd();
1935             fjy1             = _mm_setzero_pd();
1936             fjz1             = _mm_setzero_pd();
1937             fjx2             = _mm_setzero_pd();
1938             fjy2             = _mm_setzero_pd();
1939             fjz2             = _mm_setzero_pd();
1940             fjx3             = _mm_setzero_pd();
1941             fjy3             = _mm_setzero_pd();
1942             fjz3             = _mm_setzero_pd();
1943
1944             /**************************
1945              * CALCULATE INTERACTIONS *
1946              **************************/
1947
1948             r00              = _mm_mul_pd(rsq00,rinv00);
1949
1950             /* Calculate table index by multiplying r with table scale and truncate to integer */
1951             rt               = _mm_mul_pd(r00,vftabscale);
1952             vfitab           = _mm_cvttpd_epi32(rt);
1953             vfeps            = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
1954             vfitab           = _mm_slli_epi32(vfitab,3);
1955
1956             /* CUBIC SPLINE TABLE DISPERSION */
1957             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1958             F                = _mm_setzero_pd();
1959             GMX_MM_TRANSPOSE2_PD(Y,F);
1960             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1961             H                = _mm_setzero_pd();
1962             GMX_MM_TRANSPOSE2_PD(G,H);
1963             Heps             = _mm_mul_pd(vfeps,H);
1964             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1965             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1966             fvdw6            = _mm_mul_pd(c6_00,FF);
1967
1968             /* CUBIC SPLINE TABLE REPULSION */
1969             vfitab           = _mm_add_epi32(vfitab,ifour);
1970             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1971             F                = _mm_setzero_pd();
1972             GMX_MM_TRANSPOSE2_PD(Y,F);
1973             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1974             H                = _mm_setzero_pd();
1975             GMX_MM_TRANSPOSE2_PD(G,H);
1976             Heps             = _mm_mul_pd(vfeps,H);
1977             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1978             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1979             fvdw12           = _mm_mul_pd(c12_00,FF);
1980             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1981
1982             fscal            = fvdw;
1983
1984             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1985
1986             /* Calculate temporary vectorial force */
1987             tx               = _mm_mul_pd(fscal,dx00);
1988             ty               = _mm_mul_pd(fscal,dy00);
1989             tz               = _mm_mul_pd(fscal,dz00);
1990
1991             /* Update vectorial force */
1992             fix0             = _mm_add_pd(fix0,tx);
1993             fiy0             = _mm_add_pd(fiy0,ty);
1994             fiz0             = _mm_add_pd(fiz0,tz);
1995
1996             fjx0             = _mm_add_pd(fjx0,tx);
1997             fjy0             = _mm_add_pd(fjy0,ty);
1998             fjz0             = _mm_add_pd(fjz0,tz);
1999
2000             /**************************
2001              * CALCULATE INTERACTIONS *
2002              **************************/
2003
2004             if (gmx_mm_any_lt(rsq11,rcutoff2))
2005             {
2006
2007             /* REACTION-FIELD ELECTROSTATICS */
2008             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
2009
2010             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
2011
2012             fscal            = felec;
2013
2014             fscal            = _mm_and_pd(fscal,cutoff_mask);
2015
2016             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2017
2018             /* Calculate temporary vectorial force */
2019             tx               = _mm_mul_pd(fscal,dx11);
2020             ty               = _mm_mul_pd(fscal,dy11);
2021             tz               = _mm_mul_pd(fscal,dz11);
2022
2023             /* Update vectorial force */
2024             fix1             = _mm_add_pd(fix1,tx);
2025             fiy1             = _mm_add_pd(fiy1,ty);
2026             fiz1             = _mm_add_pd(fiz1,tz);
2027
2028             fjx1             = _mm_add_pd(fjx1,tx);
2029             fjy1             = _mm_add_pd(fjy1,ty);
2030             fjz1             = _mm_add_pd(fjz1,tz);
2031
2032             }
2033
2034             /**************************
2035              * CALCULATE INTERACTIONS *
2036              **************************/
2037
2038             if (gmx_mm_any_lt(rsq12,rcutoff2))
2039             {
2040
2041             /* REACTION-FIELD ELECTROSTATICS */
2042             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
2043
2044             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
2045
2046             fscal            = felec;
2047
2048             fscal            = _mm_and_pd(fscal,cutoff_mask);
2049
2050             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2051
2052             /* Calculate temporary vectorial force */
2053             tx               = _mm_mul_pd(fscal,dx12);
2054             ty               = _mm_mul_pd(fscal,dy12);
2055             tz               = _mm_mul_pd(fscal,dz12);
2056
2057             /* Update vectorial force */
2058             fix1             = _mm_add_pd(fix1,tx);
2059             fiy1             = _mm_add_pd(fiy1,ty);
2060             fiz1             = _mm_add_pd(fiz1,tz);
2061
2062             fjx2             = _mm_add_pd(fjx2,tx);
2063             fjy2             = _mm_add_pd(fjy2,ty);
2064             fjz2             = _mm_add_pd(fjz2,tz);
2065
2066             }
2067
2068             /**************************
2069              * CALCULATE INTERACTIONS *
2070              **************************/
2071
2072             if (gmx_mm_any_lt(rsq13,rcutoff2))
2073             {
2074
2075             /* REACTION-FIELD ELECTROSTATICS */
2076             felec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_mul_pd(rinv13,rinvsq13),krf2));
2077
2078             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
2079
2080             fscal            = felec;
2081
2082             fscal            = _mm_and_pd(fscal,cutoff_mask);
2083
2084             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2085
2086             /* Calculate temporary vectorial force */
2087             tx               = _mm_mul_pd(fscal,dx13);
2088             ty               = _mm_mul_pd(fscal,dy13);
2089             tz               = _mm_mul_pd(fscal,dz13);
2090
2091             /* Update vectorial force */
2092             fix1             = _mm_add_pd(fix1,tx);
2093             fiy1             = _mm_add_pd(fiy1,ty);
2094             fiz1             = _mm_add_pd(fiz1,tz);
2095
2096             fjx3             = _mm_add_pd(fjx3,tx);
2097             fjy3             = _mm_add_pd(fjy3,ty);
2098             fjz3             = _mm_add_pd(fjz3,tz);
2099
2100             }
2101
2102             /**************************
2103              * CALCULATE INTERACTIONS *
2104              **************************/
2105
2106             if (gmx_mm_any_lt(rsq21,rcutoff2))
2107             {
2108
2109             /* REACTION-FIELD ELECTROSTATICS */
2110             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
2111
2112             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
2113
2114             fscal            = felec;
2115
2116             fscal            = _mm_and_pd(fscal,cutoff_mask);
2117
2118             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2119
2120             /* Calculate temporary vectorial force */
2121             tx               = _mm_mul_pd(fscal,dx21);
2122             ty               = _mm_mul_pd(fscal,dy21);
2123             tz               = _mm_mul_pd(fscal,dz21);
2124
2125             /* Update vectorial force */
2126             fix2             = _mm_add_pd(fix2,tx);
2127             fiy2             = _mm_add_pd(fiy2,ty);
2128             fiz2             = _mm_add_pd(fiz2,tz);
2129
2130             fjx1             = _mm_add_pd(fjx1,tx);
2131             fjy1             = _mm_add_pd(fjy1,ty);
2132             fjz1             = _mm_add_pd(fjz1,tz);
2133
2134             }
2135
2136             /**************************
2137              * CALCULATE INTERACTIONS *
2138              **************************/
2139
2140             if (gmx_mm_any_lt(rsq22,rcutoff2))
2141             {
2142
2143             /* REACTION-FIELD ELECTROSTATICS */
2144             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
2145
2146             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
2147
2148             fscal            = felec;
2149
2150             fscal            = _mm_and_pd(fscal,cutoff_mask);
2151
2152             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2153
2154             /* Calculate temporary vectorial force */
2155             tx               = _mm_mul_pd(fscal,dx22);
2156             ty               = _mm_mul_pd(fscal,dy22);
2157             tz               = _mm_mul_pd(fscal,dz22);
2158
2159             /* Update vectorial force */
2160             fix2             = _mm_add_pd(fix2,tx);
2161             fiy2             = _mm_add_pd(fiy2,ty);
2162             fiz2             = _mm_add_pd(fiz2,tz);
2163
2164             fjx2             = _mm_add_pd(fjx2,tx);
2165             fjy2             = _mm_add_pd(fjy2,ty);
2166             fjz2             = _mm_add_pd(fjz2,tz);
2167
2168             }
2169
2170             /**************************
2171              * CALCULATE INTERACTIONS *
2172              **************************/
2173
2174             if (gmx_mm_any_lt(rsq23,rcutoff2))
2175             {
2176
2177             /* REACTION-FIELD ELECTROSTATICS */
2178             felec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_mul_pd(rinv23,rinvsq23),krf2));
2179
2180             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
2181
2182             fscal            = felec;
2183
2184             fscal            = _mm_and_pd(fscal,cutoff_mask);
2185
2186             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2187
2188             /* Calculate temporary vectorial force */
2189             tx               = _mm_mul_pd(fscal,dx23);
2190             ty               = _mm_mul_pd(fscal,dy23);
2191             tz               = _mm_mul_pd(fscal,dz23);
2192
2193             /* Update vectorial force */
2194             fix2             = _mm_add_pd(fix2,tx);
2195             fiy2             = _mm_add_pd(fiy2,ty);
2196             fiz2             = _mm_add_pd(fiz2,tz);
2197
2198             fjx3             = _mm_add_pd(fjx3,tx);
2199             fjy3             = _mm_add_pd(fjy3,ty);
2200             fjz3             = _mm_add_pd(fjz3,tz);
2201
2202             }
2203
2204             /**************************
2205              * CALCULATE INTERACTIONS *
2206              **************************/
2207
2208             if (gmx_mm_any_lt(rsq31,rcutoff2))
2209             {
2210
2211             /* REACTION-FIELD ELECTROSTATICS */
2212             felec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_mul_pd(rinv31,rinvsq31),krf2));
2213
2214             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
2215
2216             fscal            = felec;
2217
2218             fscal            = _mm_and_pd(fscal,cutoff_mask);
2219
2220             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2221
2222             /* Calculate temporary vectorial force */
2223             tx               = _mm_mul_pd(fscal,dx31);
2224             ty               = _mm_mul_pd(fscal,dy31);
2225             tz               = _mm_mul_pd(fscal,dz31);
2226
2227             /* Update vectorial force */
2228             fix3             = _mm_add_pd(fix3,tx);
2229             fiy3             = _mm_add_pd(fiy3,ty);
2230             fiz3             = _mm_add_pd(fiz3,tz);
2231
2232             fjx1             = _mm_add_pd(fjx1,tx);
2233             fjy1             = _mm_add_pd(fjy1,ty);
2234             fjz1             = _mm_add_pd(fjz1,tz);
2235
2236             }
2237
2238             /**************************
2239              * CALCULATE INTERACTIONS *
2240              **************************/
2241
2242             if (gmx_mm_any_lt(rsq32,rcutoff2))
2243             {
2244
2245             /* REACTION-FIELD ELECTROSTATICS */
2246             felec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_mul_pd(rinv32,rinvsq32),krf2));
2247
2248             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
2249
2250             fscal            = felec;
2251
2252             fscal            = _mm_and_pd(fscal,cutoff_mask);
2253
2254             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2255
2256             /* Calculate temporary vectorial force */
2257             tx               = _mm_mul_pd(fscal,dx32);
2258             ty               = _mm_mul_pd(fscal,dy32);
2259             tz               = _mm_mul_pd(fscal,dz32);
2260
2261             /* Update vectorial force */
2262             fix3             = _mm_add_pd(fix3,tx);
2263             fiy3             = _mm_add_pd(fiy3,ty);
2264             fiz3             = _mm_add_pd(fiz3,tz);
2265
2266             fjx2             = _mm_add_pd(fjx2,tx);
2267             fjy2             = _mm_add_pd(fjy2,ty);
2268             fjz2             = _mm_add_pd(fjz2,tz);
2269
2270             }
2271
2272             /**************************
2273              * CALCULATE INTERACTIONS *
2274              **************************/
2275
2276             if (gmx_mm_any_lt(rsq33,rcutoff2))
2277             {
2278
2279             /* REACTION-FIELD ELECTROSTATICS */
2280             felec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_mul_pd(rinv33,rinvsq33),krf2));
2281
2282             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
2283
2284             fscal            = felec;
2285
2286             fscal            = _mm_and_pd(fscal,cutoff_mask);
2287
2288             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2289
2290             /* Calculate temporary vectorial force */
2291             tx               = _mm_mul_pd(fscal,dx33);
2292             ty               = _mm_mul_pd(fscal,dy33);
2293             tz               = _mm_mul_pd(fscal,dz33);
2294
2295             /* Update vectorial force */
2296             fix3             = _mm_add_pd(fix3,tx);
2297             fiy3             = _mm_add_pd(fiy3,ty);
2298             fiz3             = _mm_add_pd(fiz3,tz);
2299
2300             fjx3             = _mm_add_pd(fjx3,tx);
2301             fjy3             = _mm_add_pd(fjy3,ty);
2302             fjz3             = _mm_add_pd(fjz3,tz);
2303
2304             }
2305
2306             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2307
2308             /* Inner loop uses 321 flops */
2309         }
2310
2311         /* End of innermost loop */
2312
2313         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2314                                               f+i_coord_offset,fshift+i_shift_offset);
2315
2316         /* Increment number of inner iterations */
2317         inneriter                  += j_index_end - j_index_start;
2318
2319         /* Outer loop uses 24 flops */
2320     }
2321
2322     /* Increment number of outer iterations */
2323     outeriter        += nri;
2324
2325     /* Update outer/inner flops */
2326
2327     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*321);
2328 }