src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_double/nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_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 "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_sse2_double.h"
  48 #include "kernelutil_x86_sse2_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_VF_sse2_double
  52  * Electrostatics interaction: ReactionField
  53  * VdW interaction:            CubicSplineTable
  54  * Geometry:                   Water3-Water3
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_VF_sse2_double
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB;
  75     int              j_coord_offsetA,j_coord_offsetB;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  80     int              vdwioffset0;
  81     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  82     int              vdwioffset1;
  83     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  84     int              vdwioffset2;
  85     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  86     int              vdwjidx0A,vdwjidx0B;
  87     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  88     int              vdwjidx1A,vdwjidx1B;
  89     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  90     int              vdwjidx2A,vdwjidx2B;
  91     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  92     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  93     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  94     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  95     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  96     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  97     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  98     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  99     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 100     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 101     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 102     real             *charge;
 103     int              nvdwtype;
 104     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 105     int              *vdwtype;
 106     real             *vdwparam;
 107     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 108     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 109     __m128i          vfitab;
 110     __m128i          ifour       = _mm_set1_epi32(4);
 111     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
 112     real             *vftab;
 113     __m128d          dummy_mask,cutoff_mask;
 114     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 115     __m128d          one     = _mm_set1_pd(1.0);
 116     __m128d          two     = _mm_set1_pd(2.0);
 117     x                = xx[0];
 118     f                = ff[0];
 119
 120     nri              = nlist->nri;
 121     iinr             = nlist->iinr;
 122     jindex           = nlist->jindex;
 123     jjnr             = nlist->jjnr;
 124     shiftidx         = nlist->shift;
 125     gid              = nlist->gid;
 126     shiftvec         = fr->shift_vec[0];
 127     fshift           = fr->fshift[0];
 128     facel            = _mm_set1_pd(fr->epsfac);
 129     charge           = mdatoms->chargeA;
 130     krf              = _mm_set1_pd(fr->ic->k_rf);
 131     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 132     crf              = _mm_set1_pd(fr->ic->c_rf);
 133     nvdwtype         = fr->ntype;
 134     vdwparam         = fr->nbfp;
 135     vdwtype          = mdatoms->typeA;
 136
 137     vftab            = kernel_data->table_vdw->data;
 138     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
 139
 140     /* Setup water-specific parameters */
 141     inr              = nlist->iinr[0];
 142     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 143     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 144     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 145     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 146
 147     jq0              = _mm_set1_pd(charge[inr+0]);
 148     jq1              = _mm_set1_pd(charge[inr+1]);
 149     jq2              = _mm_set1_pd(charge[inr+2]);
 150     vdwjidx0A        = 2*vdwtype[inr+0];
 151     qq00             = _mm_mul_pd(iq0,jq0);
 152     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 153     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 154     qq01             = _mm_mul_pd(iq0,jq1);
 155     qq02             = _mm_mul_pd(iq0,jq2);
 156     qq10             = _mm_mul_pd(iq1,jq0);
 157     qq11             = _mm_mul_pd(iq1,jq1);
 158     qq12             = _mm_mul_pd(iq1,jq2);
 159     qq20             = _mm_mul_pd(iq2,jq0);
 160     qq21             = _mm_mul_pd(iq2,jq1);
 161     qq22             = _mm_mul_pd(iq2,jq2);
 162
 163     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 164     rcutoff_scalar   = fr->rcoulomb;
 165     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 166     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 167
 168     /* Avoid stupid compiler warnings */
 169     jnrA = jnrB = 0;
 170     j_coord_offsetA = 0;
 171     j_coord_offsetB = 0;
 172
 173     outeriter        = 0;
 174     inneriter        = 0;
 175
 176     /* Start outer loop over neighborlists */
 177     for(iidx=0; iidx<nri; iidx++)
 178     {
 179         /* Load shift vector for this list */
 180         i_shift_offset   = DIM*shiftidx[iidx];
 181
 182         /* Load limits for loop over neighbors */
 183         j_index_start    = jindex[iidx];
 184         j_index_end      = jindex[iidx+1];
 185
 186         /* Get outer coordinate index */
 187         inr              = iinr[iidx];
 188         i_coord_offset   = DIM*inr;
 189
 190         /* Load i particle coords and add shift vector */
 191         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 192                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 193
 194         fix0             = _mm_setzero_pd();
 195         fiy0             = _mm_setzero_pd();
 196         fiz0             = _mm_setzero_pd();
 197         fix1             = _mm_setzero_pd();
 198         fiy1             = _mm_setzero_pd();
 199         fiz1             = _mm_setzero_pd();
 200         fix2             = _mm_setzero_pd();
 201         fiy2             = _mm_setzero_pd();
 202         fiz2             = _mm_setzero_pd();
 203
 204         /* Reset potential sums */
 205         velecsum         = _mm_setzero_pd();
 206         vvdwsum          = _mm_setzero_pd();
 207
 208         /* Start inner kernel loop */
 209         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 210         {
 211
 212             /* Get j neighbor index, and coordinate index */
 213             jnrA             = jjnr[jidx];
 214             jnrB             = jjnr[jidx+1];
 215             j_coord_offsetA  = DIM*jnrA;
 216             j_coord_offsetB  = DIM*jnrB;
 217
 218             /* load j atom coordinates */
 219             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 220                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 221
 222             /* Calculate displacement vector */
 223             dx00             = _mm_sub_pd(ix0,jx0);
 224             dy00             = _mm_sub_pd(iy0,jy0);
 225             dz00             = _mm_sub_pd(iz0,jz0);
 226             dx01             = _mm_sub_pd(ix0,jx1);
 227             dy01             = _mm_sub_pd(iy0,jy1);
 228             dz01             = _mm_sub_pd(iz0,jz1);
 229             dx02             = _mm_sub_pd(ix0,jx2);
 230             dy02             = _mm_sub_pd(iy0,jy2);
 231             dz02             = _mm_sub_pd(iz0,jz2);
 232             dx10             = _mm_sub_pd(ix1,jx0);
 233             dy10             = _mm_sub_pd(iy1,jy0);
 234             dz10             = _mm_sub_pd(iz1,jz0);
 235             dx11             = _mm_sub_pd(ix1,jx1);
 236             dy11             = _mm_sub_pd(iy1,jy1);
 237             dz11             = _mm_sub_pd(iz1,jz1);
 238             dx12             = _mm_sub_pd(ix1,jx2);
 239             dy12             = _mm_sub_pd(iy1,jy2);
 240             dz12             = _mm_sub_pd(iz1,jz2);
 241             dx20             = _mm_sub_pd(ix2,jx0);
 242             dy20             = _mm_sub_pd(iy2,jy0);
 243             dz20             = _mm_sub_pd(iz2,jz0);
 244             dx21             = _mm_sub_pd(ix2,jx1);
 245             dy21             = _mm_sub_pd(iy2,jy1);
 246             dz21             = _mm_sub_pd(iz2,jz1);
 247             dx22             = _mm_sub_pd(ix2,jx2);
 248             dy22             = _mm_sub_pd(iy2,jy2);
 249             dz22             = _mm_sub_pd(iz2,jz2);
 250
 251             /* Calculate squared distance and things based on it */
 252             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 253             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 254             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 255             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 256             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 257             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 258             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 259             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 260             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 261
 262             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 263             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 264             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 265             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 266             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 267             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 268             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 269             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 270             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 271
 272             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 273             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 274             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 275             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 276             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 277             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 278             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 279             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 280             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 281
 282             fjx0             = _mm_setzero_pd();
 283             fjy0             = _mm_setzero_pd();
 284             fjz0             = _mm_setzero_pd();
 285             fjx1             = _mm_setzero_pd();
 286             fjy1             = _mm_setzero_pd();
 287             fjz1             = _mm_setzero_pd();
 288             fjx2             = _mm_setzero_pd();
 289             fjy2             = _mm_setzero_pd();
 290             fjz2             = _mm_setzero_pd();
 291
 292             /**************************
 293              * CALCULATE INTERACTIONS *
 294              **************************/
 295
 296             if (gmx_mm_any_lt(rsq00,rcutoff2))
 297             {
 298
 299             r00              = _mm_mul_pd(rsq00,rinv00);
 300
 301             /* Calculate table index by multiplying r with table scale and truncate to integer */
 302             rt               = _mm_mul_pd(r00,vftabscale);
 303             vfitab           = _mm_cvttpd_epi32(rt);
 304             vfeps            = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
 305             vfitab           = _mm_slli_epi32(vfitab,3);
 306
 307             /* REACTION-FIELD ELECTROSTATICS */
 308             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
 309             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 310
 311             /* CUBIC SPLINE TABLE DISPERSION */
 312             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 313             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
 314             GMX_MM_TRANSPOSE2_PD(Y,F);
 315             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 316             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
 317             GMX_MM_TRANSPOSE2_PD(G,H);
 318             Heps             = _mm_mul_pd(vfeps,H);
 319             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 320             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 321             vvdw6            = _mm_mul_pd(c6_00,VV);
 322             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 323             fvdw6            = _mm_mul_pd(c6_00,FF);
 324
 325             /* CUBIC SPLINE TABLE REPULSION */
 326             vfitab           = _mm_add_epi32(vfitab,ifour);
 327             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 328             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
 329             GMX_MM_TRANSPOSE2_PD(Y,F);
 330             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 331             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
 332             GMX_MM_TRANSPOSE2_PD(G,H);
 333             Heps             = _mm_mul_pd(vfeps,H);
 334             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 335             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 336             vvdw12           = _mm_mul_pd(c12_00,VV);
 337             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 338             fvdw12           = _mm_mul_pd(c12_00,FF);
 339             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 340             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 341
 342             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 343
 344             /* Update potential sum for this i atom from the interaction with this j atom. */
 345             velec            = _mm_and_pd(velec,cutoff_mask);
 346             velecsum         = _mm_add_pd(velecsum,velec);
 347             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 348             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 349
 350             fscal            = _mm_add_pd(felec,fvdw);
 351
 352             fscal            = _mm_and_pd(fscal,cutoff_mask);
 353
 354             /* Calculate temporary vectorial force */
 355             tx               = _mm_mul_pd(fscal,dx00);
 356             ty               = _mm_mul_pd(fscal,dy00);
 357             tz               = _mm_mul_pd(fscal,dz00);
 358
 359             /* Update vectorial force */
 360             fix0             = _mm_add_pd(fix0,tx);
 361             fiy0             = _mm_add_pd(fiy0,ty);
 362             fiz0             = _mm_add_pd(fiz0,tz);
 363
 364             fjx0             = _mm_add_pd(fjx0,tx);
 365             fjy0             = _mm_add_pd(fjy0,ty);
 366             fjz0             = _mm_add_pd(fjz0,tz);
 367
 368             }
 369
 370             /**************************
 371              * CALCULATE INTERACTIONS *
 372              **************************/
 373
 374             if (gmx_mm_any_lt(rsq01,rcutoff2))
 375             {
 376
 377             /* REACTION-FIELD ELECTROSTATICS */
 378             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_add_pd(rinv01,_mm_mul_pd(krf,rsq01)),crf));
 379             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
 380
 381             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
 382
 383             /* Update potential sum for this i atom from the interaction with this j atom. */
 384             velec            = _mm_and_pd(velec,cutoff_mask);
 385             velecsum         = _mm_add_pd(velecsum,velec);
 386
 387             fscal            = felec;
 388
 389             fscal            = _mm_and_pd(fscal,cutoff_mask);
 390
 391             /* Calculate temporary vectorial force */
 392             tx               = _mm_mul_pd(fscal,dx01);
 393             ty               = _mm_mul_pd(fscal,dy01);
 394             tz               = _mm_mul_pd(fscal,dz01);
 395
 396             /* Update vectorial force */
 397             fix0             = _mm_add_pd(fix0,tx);
 398             fiy0             = _mm_add_pd(fiy0,ty);
 399             fiz0             = _mm_add_pd(fiz0,tz);
 400
 401             fjx1             = _mm_add_pd(fjx1,tx);
 402             fjy1             = _mm_add_pd(fjy1,ty);
 403             fjz1             = _mm_add_pd(fjz1,tz);
 404
 405             }
 406
 407             /**************************
 408              * CALCULATE INTERACTIONS *
 409              **************************/
 410
 411             if (gmx_mm_any_lt(rsq02,rcutoff2))
 412             {
 413
 414             /* REACTION-FIELD ELECTROSTATICS */
 415             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_add_pd(rinv02,_mm_mul_pd(krf,rsq02)),crf));
 416             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
 417
 418             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
 419
 420             /* Update potential sum for this i atom from the interaction with this j atom. */
 421             velec            = _mm_and_pd(velec,cutoff_mask);
 422             velecsum         = _mm_add_pd(velecsum,velec);
 423
 424             fscal            = felec;
 425
 426             fscal            = _mm_and_pd(fscal,cutoff_mask);
 427
 428             /* Calculate temporary vectorial force */
 429             tx               = _mm_mul_pd(fscal,dx02);
 430             ty               = _mm_mul_pd(fscal,dy02);
 431             tz               = _mm_mul_pd(fscal,dz02);
 432
 433             /* Update vectorial force */
 434             fix0             = _mm_add_pd(fix0,tx);
 435             fiy0             = _mm_add_pd(fiy0,ty);
 436             fiz0             = _mm_add_pd(fiz0,tz);
 437
 438             fjx2             = _mm_add_pd(fjx2,tx);
 439             fjy2             = _mm_add_pd(fjy2,ty);
 440             fjz2             = _mm_add_pd(fjz2,tz);
 441
 442             }
 443
 444             /**************************
 445              * CALCULATE INTERACTIONS *
 446              **************************/
 447
 448             if (gmx_mm_any_lt(rsq10,rcutoff2))
 449             {
 450
 451             /* REACTION-FIELD ELECTROSTATICS */
 452             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 453             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 454
 455             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
 456
 457             /* Update potential sum for this i atom from the interaction with this j atom. */
 458             velec            = _mm_and_pd(velec,cutoff_mask);
 459             velecsum         = _mm_add_pd(velecsum,velec);
 460
 461             fscal            = felec;
 462
 463             fscal            = _mm_and_pd(fscal,cutoff_mask);
 464
 465             /* Calculate temporary vectorial force */
 466             tx               = _mm_mul_pd(fscal,dx10);
 467             ty               = _mm_mul_pd(fscal,dy10);
 468             tz               = _mm_mul_pd(fscal,dz10);
 469
 470             /* Update vectorial force */
 471             fix1             = _mm_add_pd(fix1,tx);
 472             fiy1             = _mm_add_pd(fiy1,ty);
 473             fiz1             = _mm_add_pd(fiz1,tz);
 474
 475             fjx0             = _mm_add_pd(fjx0,tx);
 476             fjy0             = _mm_add_pd(fjy0,ty);
 477             fjz0             = _mm_add_pd(fjz0,tz);
 478
 479             }
 480
 481             /**************************
 482              * CALCULATE INTERACTIONS *
 483              **************************/
 484
 485             if (gmx_mm_any_lt(rsq11,rcutoff2))
 486             {
 487
 488             /* REACTION-FIELD ELECTROSTATICS */
 489             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 490             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 491
 492             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 493
 494             /* Update potential sum for this i atom from the interaction with this j atom. */
 495             velec            = _mm_and_pd(velec,cutoff_mask);
 496             velecsum         = _mm_add_pd(velecsum,velec);
 497
 498             fscal            = felec;
 499
 500             fscal            = _mm_and_pd(fscal,cutoff_mask);
 501
 502             /* Calculate temporary vectorial force */
 503             tx               = _mm_mul_pd(fscal,dx11);
 504             ty               = _mm_mul_pd(fscal,dy11);
 505             tz               = _mm_mul_pd(fscal,dz11);
 506
 507             /* Update vectorial force */
 508             fix1             = _mm_add_pd(fix1,tx);
 509             fiy1             = _mm_add_pd(fiy1,ty);
 510             fiz1             = _mm_add_pd(fiz1,tz);
 511
 512             fjx1             = _mm_add_pd(fjx1,tx);
 513             fjy1             = _mm_add_pd(fjy1,ty);
 514             fjz1             = _mm_add_pd(fjz1,tz);
 515
 516             }
 517
 518             /**************************
 519              * CALCULATE INTERACTIONS *
 520              **************************/
 521
 522             if (gmx_mm_any_lt(rsq12,rcutoff2))
 523             {
 524
 525             /* REACTION-FIELD ELECTROSTATICS */
 526             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 527             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 528
 529             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 530
 531             /* Update potential sum for this i atom from the interaction with this j atom. */
 532             velec            = _mm_and_pd(velec,cutoff_mask);
 533             velecsum         = _mm_add_pd(velecsum,velec);
 534
 535             fscal            = felec;
 536
 537             fscal            = _mm_and_pd(fscal,cutoff_mask);
 538
 539             /* Calculate temporary vectorial force */
 540             tx               = _mm_mul_pd(fscal,dx12);
 541             ty               = _mm_mul_pd(fscal,dy12);
 542             tz               = _mm_mul_pd(fscal,dz12);
 543
 544             /* Update vectorial force */
 545             fix1             = _mm_add_pd(fix1,tx);
 546             fiy1             = _mm_add_pd(fiy1,ty);
 547             fiz1             = _mm_add_pd(fiz1,tz);
 548
 549             fjx2             = _mm_add_pd(fjx2,tx);
 550             fjy2             = _mm_add_pd(fjy2,ty);
 551             fjz2             = _mm_add_pd(fjz2,tz);
 552
 553             }
 554
 555             /**************************
 556              * CALCULATE INTERACTIONS *
 557              **************************/
 558
 559             if (gmx_mm_any_lt(rsq20,rcutoff2))
 560             {
 561
 562             /* REACTION-FIELD ELECTROSTATICS */
 563             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 564             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 565
 566             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
 567
 568             /* Update potential sum for this i atom from the interaction with this j atom. */
 569             velec            = _mm_and_pd(velec,cutoff_mask);
 570             velecsum         = _mm_add_pd(velecsum,velec);
 571
 572             fscal            = felec;
 573
 574             fscal            = _mm_and_pd(fscal,cutoff_mask);
 575
 576             /* Calculate temporary vectorial force */
 577             tx               = _mm_mul_pd(fscal,dx20);
 578             ty               = _mm_mul_pd(fscal,dy20);
 579             tz               = _mm_mul_pd(fscal,dz20);
 580
 581             /* Update vectorial force */
 582             fix2             = _mm_add_pd(fix2,tx);
 583             fiy2             = _mm_add_pd(fiy2,ty);
 584             fiz2             = _mm_add_pd(fiz2,tz);
 585
 586             fjx0             = _mm_add_pd(fjx0,tx);
 587             fjy0             = _mm_add_pd(fjy0,ty);
 588             fjz0             = _mm_add_pd(fjz0,tz);
 589
 590             }
 591
 592             /**************************
 593              * CALCULATE INTERACTIONS *
 594              **************************/
 595
 596             if (gmx_mm_any_lt(rsq21,rcutoff2))
 597             {
 598
 599             /* REACTION-FIELD ELECTROSTATICS */
 600             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 601             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 602
 603             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 604
 605             /* Update potential sum for this i atom from the interaction with this j atom. */
 606             velec            = _mm_and_pd(velec,cutoff_mask);
 607             velecsum         = _mm_add_pd(velecsum,velec);
 608
 609             fscal            = felec;
 610
 611             fscal            = _mm_and_pd(fscal,cutoff_mask);
 612
 613             /* Calculate temporary vectorial force */
 614             tx               = _mm_mul_pd(fscal,dx21);
 615             ty               = _mm_mul_pd(fscal,dy21);
 616             tz               = _mm_mul_pd(fscal,dz21);
 617
 618             /* Update vectorial force */
 619             fix2             = _mm_add_pd(fix2,tx);
 620             fiy2             = _mm_add_pd(fiy2,ty);
 621             fiz2             = _mm_add_pd(fiz2,tz);
 622
 623             fjx1             = _mm_add_pd(fjx1,tx);
 624             fjy1             = _mm_add_pd(fjy1,ty);
 625             fjz1             = _mm_add_pd(fjz1,tz);
 626
 627             }
 628
 629             /**************************
 630              * CALCULATE INTERACTIONS *
 631              **************************/
 632
 633             if (gmx_mm_any_lt(rsq22,rcutoff2))
 634             {
 635
 636             /* REACTION-FIELD ELECTROSTATICS */
 637             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 638             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 639
 640             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 641
 642             /* Update potential sum for this i atom from the interaction with this j atom. */
 643             velec            = _mm_and_pd(velec,cutoff_mask);
 644             velecsum         = _mm_add_pd(velecsum,velec);
 645
 646             fscal            = felec;
 647
 648             fscal            = _mm_and_pd(fscal,cutoff_mask);
 649
 650             /* Calculate temporary vectorial force */
 651             tx               = _mm_mul_pd(fscal,dx22);
 652             ty               = _mm_mul_pd(fscal,dy22);
 653             tz               = _mm_mul_pd(fscal,dz22);
 654
 655             /* Update vectorial force */
 656             fix2             = _mm_add_pd(fix2,tx);
 657             fiy2             = _mm_add_pd(fiy2,ty);
 658             fiz2             = _mm_add_pd(fiz2,tz);
 659
 660             fjx2             = _mm_add_pd(fjx2,tx);
 661             fjy2             = _mm_add_pd(fjy2,ty);
 662             fjz2             = _mm_add_pd(fjz2,tz);
 663
 664             }
 665
 666             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 667
 668             /* Inner loop uses 360 flops */
 669         }
 670
 671         if(jidx<j_index_end)
 672         {
 673
 674             jnrA             = jjnr[jidx];
 675             j_coord_offsetA  = DIM*jnrA;
 676
 677             /* load j atom coordinates */
 678             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 679                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 680
 681             /* Calculate displacement vector */
 682             dx00             = _mm_sub_pd(ix0,jx0);
 683             dy00             = _mm_sub_pd(iy0,jy0);
 684             dz00             = _mm_sub_pd(iz0,jz0);
 685             dx01             = _mm_sub_pd(ix0,jx1);
 686             dy01             = _mm_sub_pd(iy0,jy1);
 687             dz01             = _mm_sub_pd(iz0,jz1);
 688             dx02             = _mm_sub_pd(ix0,jx2);
 689             dy02             = _mm_sub_pd(iy0,jy2);
 690             dz02             = _mm_sub_pd(iz0,jz2);
 691             dx10             = _mm_sub_pd(ix1,jx0);
 692             dy10             = _mm_sub_pd(iy1,jy0);
 693             dz10             = _mm_sub_pd(iz1,jz0);
 694             dx11             = _mm_sub_pd(ix1,jx1);
 695             dy11             = _mm_sub_pd(iy1,jy1);
 696             dz11             = _mm_sub_pd(iz1,jz1);
 697             dx12             = _mm_sub_pd(ix1,jx2);
 698             dy12             = _mm_sub_pd(iy1,jy2);
 699             dz12             = _mm_sub_pd(iz1,jz2);
 700             dx20             = _mm_sub_pd(ix2,jx0);
 701             dy20             = _mm_sub_pd(iy2,jy0);
 702             dz20             = _mm_sub_pd(iz2,jz0);
 703             dx21             = _mm_sub_pd(ix2,jx1);
 704             dy21             = _mm_sub_pd(iy2,jy1);
 705             dz21             = _mm_sub_pd(iz2,jz1);
 706             dx22             = _mm_sub_pd(ix2,jx2);
 707             dy22             = _mm_sub_pd(iy2,jy2);
 708             dz22             = _mm_sub_pd(iz2,jz2);
 709
 710             /* Calculate squared distance and things based on it */
 711             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 712             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 713             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 714             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 715             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 716             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 717             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 718             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 719             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 720
 721             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 722             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 723             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 724             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 725             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 726             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 727             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 728             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 729             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 730
 731             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 732             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 733             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 734             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 735             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 736             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 737             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 738             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 739             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 740
 741             fjx0             = _mm_setzero_pd();
 742             fjy0             = _mm_setzero_pd();
 743             fjz0             = _mm_setzero_pd();
 744             fjx1             = _mm_setzero_pd();
 745             fjy1             = _mm_setzero_pd();
 746             fjz1             = _mm_setzero_pd();
 747             fjx2             = _mm_setzero_pd();
 748             fjy2             = _mm_setzero_pd();
 749             fjz2             = _mm_setzero_pd();
 750
 751             /**************************
 752              * CALCULATE INTERACTIONS *
 753              **************************/
 754
 755             if (gmx_mm_any_lt(rsq00,rcutoff2))
 756             {
 757
 758             r00              = _mm_mul_pd(rsq00,rinv00);
 759
 760             /* Calculate table index by multiplying r with table scale and truncate to integer */
 761             rt               = _mm_mul_pd(r00,vftabscale);
 762             vfitab           = _mm_cvttpd_epi32(rt);
 763             vfeps            = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
 764             vfitab           = _mm_slli_epi32(vfitab,3);
 765
 766             /* REACTION-FIELD ELECTROSTATICS */
 767             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
 768             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 769
 770             /* CUBIC SPLINE TABLE DISPERSION */
 771             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 772             F                = _mm_setzero_pd();
 773             GMX_MM_TRANSPOSE2_PD(Y,F);
 774             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 775             H                = _mm_setzero_pd();
 776             GMX_MM_TRANSPOSE2_PD(G,H);
 777             Heps             = _mm_mul_pd(vfeps,H);
 778             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 779             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 780             vvdw6            = _mm_mul_pd(c6_00,VV);
 781             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 782             fvdw6            = _mm_mul_pd(c6_00,FF);
 783
 784             /* CUBIC SPLINE TABLE REPULSION */
 785             vfitab           = _mm_add_epi32(vfitab,ifour);
 786             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 787             F                = _mm_setzero_pd();
 788             GMX_MM_TRANSPOSE2_PD(Y,F);
 789             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 790             H                = _mm_setzero_pd();
 791             GMX_MM_TRANSPOSE2_PD(G,H);
 792             Heps             = _mm_mul_pd(vfeps,H);
 793             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 794             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 795             vvdw12           = _mm_mul_pd(c12_00,VV);
 796             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 797             fvdw12           = _mm_mul_pd(c12_00,FF);
 798             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 799             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 800
 801             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 802
 803             /* Update potential sum for this i atom from the interaction with this j atom. */
 804             velec            = _mm_and_pd(velec,cutoff_mask);
 805             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 806             velecsum         = _mm_add_pd(velecsum,velec);
 807             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 808             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 809             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 810
 811             fscal            = _mm_add_pd(felec,fvdw);
 812
 813             fscal            = _mm_and_pd(fscal,cutoff_mask);
 814
 815             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 816
 817             /* Calculate temporary vectorial force */
 818             tx               = _mm_mul_pd(fscal,dx00);
 819             ty               = _mm_mul_pd(fscal,dy00);
 820             tz               = _mm_mul_pd(fscal,dz00);
 821
 822             /* Update vectorial force */
 823             fix0             = _mm_add_pd(fix0,tx);
 824             fiy0             = _mm_add_pd(fiy0,ty);
 825             fiz0             = _mm_add_pd(fiz0,tz);
 826
 827             fjx0             = _mm_add_pd(fjx0,tx);
 828             fjy0             = _mm_add_pd(fjy0,ty);
 829             fjz0             = _mm_add_pd(fjz0,tz);
 830
 831             }
 832
 833             /**************************
 834              * CALCULATE INTERACTIONS *
 835              **************************/
 836
 837             if (gmx_mm_any_lt(rsq01,rcutoff2))
 838             {
 839
 840             /* REACTION-FIELD ELECTROSTATICS */
 841             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_add_pd(rinv01,_mm_mul_pd(krf,rsq01)),crf));
 842             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
 843
 844             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
 845
 846             /* Update potential sum for this i atom from the interaction with this j atom. */
 847             velec            = _mm_and_pd(velec,cutoff_mask);
 848             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 849             velecsum         = _mm_add_pd(velecsum,velec);
 850
 851             fscal            = felec;
 852
 853             fscal            = _mm_and_pd(fscal,cutoff_mask);
 854
 855             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 856
 857             /* Calculate temporary vectorial force */
 858             tx               = _mm_mul_pd(fscal,dx01);
 859             ty               = _mm_mul_pd(fscal,dy01);
 860             tz               = _mm_mul_pd(fscal,dz01);
 861
 862             /* Update vectorial force */
 863             fix0             = _mm_add_pd(fix0,tx);
 864             fiy0             = _mm_add_pd(fiy0,ty);
 865             fiz0             = _mm_add_pd(fiz0,tz);
 866
 867             fjx1             = _mm_add_pd(fjx1,tx);
 868             fjy1             = _mm_add_pd(fjy1,ty);
 869             fjz1             = _mm_add_pd(fjz1,tz);
 870
 871             }
 872
 873             /**************************
 874              * CALCULATE INTERACTIONS *
 875              **************************/
 876
 877             if (gmx_mm_any_lt(rsq02,rcutoff2))
 878             {
 879
 880             /* REACTION-FIELD ELECTROSTATICS */
 881             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_add_pd(rinv02,_mm_mul_pd(krf,rsq02)),crf));
 882             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
 883
 884             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
 885
 886             /* Update potential sum for this i atom from the interaction with this j atom. */
 887             velec            = _mm_and_pd(velec,cutoff_mask);
 888             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 889             velecsum         = _mm_add_pd(velecsum,velec);
 890
 891             fscal            = felec;
 892
 893             fscal            = _mm_and_pd(fscal,cutoff_mask);
 894
 895             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 896
 897             /* Calculate temporary vectorial force */
 898             tx               = _mm_mul_pd(fscal,dx02);
 899             ty               = _mm_mul_pd(fscal,dy02);
 900             tz               = _mm_mul_pd(fscal,dz02);
 901
 902             /* Update vectorial force */
 903             fix0             = _mm_add_pd(fix0,tx);
 904             fiy0             = _mm_add_pd(fiy0,ty);
 905             fiz0             = _mm_add_pd(fiz0,tz);
 906
 907             fjx2             = _mm_add_pd(fjx2,tx);
 908             fjy2             = _mm_add_pd(fjy2,ty);
 909             fjz2             = _mm_add_pd(fjz2,tz);
 910
 911             }
 912
 913             /**************************
 914              * CALCULATE INTERACTIONS *
 915              **************************/
 916
 917             if (gmx_mm_any_lt(rsq10,rcutoff2))
 918             {
 919
 920             /* REACTION-FIELD ELECTROSTATICS */
 921             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 922             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 923
 924             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
 925
 926             /* Update potential sum for this i atom from the interaction with this j atom. */
 927             velec            = _mm_and_pd(velec,cutoff_mask);
 928             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 929             velecsum         = _mm_add_pd(velecsum,velec);
 930
 931             fscal            = felec;
 932
 933             fscal            = _mm_and_pd(fscal,cutoff_mask);
 934
 935             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 936
 937             /* Calculate temporary vectorial force */
 938             tx               = _mm_mul_pd(fscal,dx10);
 939             ty               = _mm_mul_pd(fscal,dy10);
 940             tz               = _mm_mul_pd(fscal,dz10);
 941
 942             /* Update vectorial force */
 943             fix1             = _mm_add_pd(fix1,tx);
 944             fiy1             = _mm_add_pd(fiy1,ty);
 945             fiz1             = _mm_add_pd(fiz1,tz);
 946
 947             fjx0             = _mm_add_pd(fjx0,tx);
 948             fjy0             = _mm_add_pd(fjy0,ty);
 949             fjz0             = _mm_add_pd(fjz0,tz);
 950
 951             }
 952
 953             /**************************
 954              * CALCULATE INTERACTIONS *
 955              **************************/
 956
 957             if (gmx_mm_any_lt(rsq11,rcutoff2))
 958             {
 959
 960             /* REACTION-FIELD ELECTROSTATICS */
 961             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 962             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 963
 964             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 965
 966             /* Update potential sum for this i atom from the interaction with this j atom. */
 967             velec            = _mm_and_pd(velec,cutoff_mask);
 968             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 969             velecsum         = _mm_add_pd(velecsum,velec);
 970
 971             fscal            = felec;
 972
 973             fscal            = _mm_and_pd(fscal,cutoff_mask);
 974
 975             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 976
 977             /* Calculate temporary vectorial force */
 978             tx               = _mm_mul_pd(fscal,dx11);
 979             ty               = _mm_mul_pd(fscal,dy11);
 980             tz               = _mm_mul_pd(fscal,dz11);
 981
 982             /* Update vectorial force */
 983             fix1             = _mm_add_pd(fix1,tx);
 984             fiy1             = _mm_add_pd(fiy1,ty);
 985             fiz1             = _mm_add_pd(fiz1,tz);
 986
 987             fjx1             = _mm_add_pd(fjx1,tx);
 988             fjy1             = _mm_add_pd(fjy1,ty);
 989             fjz1             = _mm_add_pd(fjz1,tz);
 990
 991             }
 992
 993             /**************************
 994              * CALCULATE INTERACTIONS *
 995              **************************/
 996
 997             if (gmx_mm_any_lt(rsq12,rcutoff2))
 998             {
 999
1000             /* REACTION-FIELD ELECTROSTATICS */
1001             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
1002             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1003
1004             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1005
1006             /* Update potential sum for this i atom from the interaction with this j atom. */
1007             velec            = _mm_and_pd(velec,cutoff_mask);
1008             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1009             velecsum         = _mm_add_pd(velecsum,velec);
1010
1011             fscal            = felec;
1012
1013             fscal            = _mm_and_pd(fscal,cutoff_mask);
1014
1015             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1016
1017             /* Calculate temporary vectorial force */
1018             tx               = _mm_mul_pd(fscal,dx12);
1019             ty               = _mm_mul_pd(fscal,dy12);
1020             tz               = _mm_mul_pd(fscal,dz12);
1021
1022             /* Update vectorial force */
1023             fix1             = _mm_add_pd(fix1,tx);
1024             fiy1             = _mm_add_pd(fiy1,ty);
1025             fiz1             = _mm_add_pd(fiz1,tz);
1026
1027             fjx2             = _mm_add_pd(fjx2,tx);
1028             fjy2             = _mm_add_pd(fjy2,ty);
1029             fjz2             = _mm_add_pd(fjz2,tz);
1030
1031             }
1032
1033             /**************************
1034              * CALCULATE INTERACTIONS *
1035              **************************/
1036
1037             if (gmx_mm_any_lt(rsq20,rcutoff2))
1038             {
1039
1040             /* REACTION-FIELD ELECTROSTATICS */
1041             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
1042             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
1043
1044             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
1045
1046             /* Update potential sum for this i atom from the interaction with this j atom. */
1047             velec            = _mm_and_pd(velec,cutoff_mask);
1048             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1049             velecsum         = _mm_add_pd(velecsum,velec);
1050
1051             fscal            = felec;
1052
1053             fscal            = _mm_and_pd(fscal,cutoff_mask);
1054
1055             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1056
1057             /* Calculate temporary vectorial force */
1058             tx               = _mm_mul_pd(fscal,dx20);
1059             ty               = _mm_mul_pd(fscal,dy20);
1060             tz               = _mm_mul_pd(fscal,dz20);
1061
1062             /* Update vectorial force */
1063             fix2             = _mm_add_pd(fix2,tx);
1064             fiy2             = _mm_add_pd(fiy2,ty);
1065             fiz2             = _mm_add_pd(fiz2,tz);
1066
1067             fjx0             = _mm_add_pd(fjx0,tx);
1068             fjy0             = _mm_add_pd(fjy0,ty);
1069             fjz0             = _mm_add_pd(fjz0,tz);
1070
1071             }
1072
1073             /**************************
1074              * CALCULATE INTERACTIONS *
1075              **************************/
1076
1077             if (gmx_mm_any_lt(rsq21,rcutoff2))
1078             {
1079
1080             /* REACTION-FIELD ELECTROSTATICS */
1081             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
1082             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1083
1084             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1085
1086             /* Update potential sum for this i atom from the interaction with this j atom. */
1087             velec            = _mm_and_pd(velec,cutoff_mask);
1088             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1089             velecsum         = _mm_add_pd(velecsum,velec);
1090
1091             fscal            = felec;
1092
1093             fscal            = _mm_and_pd(fscal,cutoff_mask);
1094
1095             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1096
1097             /* Calculate temporary vectorial force */
1098             tx               = _mm_mul_pd(fscal,dx21);
1099             ty               = _mm_mul_pd(fscal,dy21);
1100             tz               = _mm_mul_pd(fscal,dz21);
1101
1102             /* Update vectorial force */
1103             fix2             = _mm_add_pd(fix2,tx);
1104             fiy2             = _mm_add_pd(fiy2,ty);
1105             fiz2             = _mm_add_pd(fiz2,tz);
1106
1107             fjx1             = _mm_add_pd(fjx1,tx);
1108             fjy1             = _mm_add_pd(fjy1,ty);
1109             fjz1             = _mm_add_pd(fjz1,tz);
1110
1111             }
1112
1113             /**************************
1114              * CALCULATE INTERACTIONS *
1115              **************************/
1116
1117             if (gmx_mm_any_lt(rsq22,rcutoff2))
1118             {
1119
1120             /* REACTION-FIELD ELECTROSTATICS */
1121             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
1122             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1123
1124             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1125
1126             /* Update potential sum for this i atom from the interaction with this j atom. */
1127             velec            = _mm_and_pd(velec,cutoff_mask);
1128             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1129             velecsum         = _mm_add_pd(velecsum,velec);
1130
1131             fscal            = felec;
1132
1133             fscal            = _mm_and_pd(fscal,cutoff_mask);
1134
1135             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1136
1137             /* Calculate temporary vectorial force */
1138             tx               = _mm_mul_pd(fscal,dx22);
1139             ty               = _mm_mul_pd(fscal,dy22);
1140             tz               = _mm_mul_pd(fscal,dz22);
1141
1142             /* Update vectorial force */
1143             fix2             = _mm_add_pd(fix2,tx);
1144             fiy2             = _mm_add_pd(fiy2,ty);
1145             fiz2             = _mm_add_pd(fiz2,tz);
1146
1147             fjx2             = _mm_add_pd(fjx2,tx);
1148             fjy2             = _mm_add_pd(fjy2,ty);
1149             fjz2             = _mm_add_pd(fjz2,tz);
1150
1151             }
1152
1153             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1154
1155             /* Inner loop uses 360 flops */
1156         }
1157
1158         /* End of innermost loop */
1159
1160         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1161                                               f+i_coord_offset,fshift+i_shift_offset);
1162
1163         ggid                        = gid[iidx];
1164         /* Update potential energies */
1165         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1166         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
1167
1168         /* Increment number of inner iterations */
1169         inneriter                  += j_index_end - j_index_start;
1170
1171         /* Outer loop uses 20 flops */
1172     }
1173
1174     /* Increment number of outer iterations */
1175     outeriter        += nri;
1176
1177     /* Update outer/inner flops */
1178
1179     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*360);
1180 }
1181 /*
1182  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_F_sse2_double
1183  * Electrostatics interaction: ReactionField
1184  * VdW interaction:            CubicSplineTable
1185  * Geometry:                   Water3-Water3
1186  * Calculate force/pot:        Force
1187  */
1188 void
1189 nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_F_sse2_double
1190                     (t_nblist                    * gmx_restrict       nlist,
1191                      rvec                        * gmx_restrict          xx,
1192                      rvec                        * gmx_restrict          ff,
1193                      t_forcerec                  * gmx_restrict          fr,
1194                      t_mdatoms                   * gmx_restrict     mdatoms,
1195                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1196                      t_nrnb                      * gmx_restrict        nrnb)
1197 {
1198     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1199      * just 0 for non-waters.
1200      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1201      * jnr indices corresponding to data put in the four positions in the SIMD register.
1202      */
1203     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1204     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1205     int              jnrA,jnrB;
1206     int              j_coord_offsetA,j_coord_offsetB;
1207     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1208     real             rcutoff_scalar;
1209     real             *shiftvec,*fshift,*x,*f;
1210     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1211     int              vdwioffset0;
1212     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1213     int              vdwioffset1;
1214     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1215     int              vdwioffset2;
1216     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1217     int              vdwjidx0A,vdwjidx0B;
1218     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1219     int              vdwjidx1A,vdwjidx1B;
1220     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1221     int              vdwjidx2A,vdwjidx2B;
1222     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1223     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1224     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1225     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1226     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1227     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1228     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1229     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1230     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1231     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1232     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1233     real             *charge;
1234     int              nvdwtype;
1235     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1236     int              *vdwtype;
1237     real             *vdwparam;
1238     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1239     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1240     __m128i          vfitab;
1241     __m128i          ifour       = _mm_set1_epi32(4);
1242     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
1243     real             *vftab;
1244     __m128d          dummy_mask,cutoff_mask;
1245     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1246     __m128d          one     = _mm_set1_pd(1.0);
1247     __m128d          two     = _mm_set1_pd(2.0);
1248     x                = xx[0];
1249     f                = ff[0];
1250
1251     nri              = nlist->nri;
1252     iinr             = nlist->iinr;
1253     jindex           = nlist->jindex;
1254     jjnr             = nlist->jjnr;
1255     shiftidx         = nlist->shift;
1256     gid              = nlist->gid;
1257     shiftvec         = fr->shift_vec[0];
1258     fshift           = fr->fshift[0];
1259     facel            = _mm_set1_pd(fr->epsfac);
1260     charge           = mdatoms->chargeA;
1261     krf              = _mm_set1_pd(fr->ic->k_rf);
1262     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1263     crf              = _mm_set1_pd(fr->ic->c_rf);
1264     nvdwtype         = fr->ntype;
1265     vdwparam         = fr->nbfp;
1266     vdwtype          = mdatoms->typeA;
1267
1268     vftab            = kernel_data->table_vdw->data;
1269     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
1270
1271     /* Setup water-specific parameters */
1272     inr              = nlist->iinr[0];
1273     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
1274     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1275     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1276     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1277
1278     jq0              = _mm_set1_pd(charge[inr+0]);
1279     jq1              = _mm_set1_pd(charge[inr+1]);
1280     jq2              = _mm_set1_pd(charge[inr+2]);
1281     vdwjidx0A        = 2*vdwtype[inr+0];
1282     qq00             = _mm_mul_pd(iq0,jq0);
1283     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1284     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1285     qq01             = _mm_mul_pd(iq0,jq1);
1286     qq02             = _mm_mul_pd(iq0,jq2);
1287     qq10             = _mm_mul_pd(iq1,jq0);
1288     qq11             = _mm_mul_pd(iq1,jq1);
1289     qq12             = _mm_mul_pd(iq1,jq2);
1290     qq20             = _mm_mul_pd(iq2,jq0);
1291     qq21             = _mm_mul_pd(iq2,jq1);
1292     qq22             = _mm_mul_pd(iq2,jq2);
1293
1294     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1295     rcutoff_scalar   = fr->rcoulomb;
1296     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1297     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1298
1299     /* Avoid stupid compiler warnings */
1300     jnrA = jnrB = 0;
1301     j_coord_offsetA = 0;
1302     j_coord_offsetB = 0;
1303
1304     outeriter        = 0;
1305     inneriter        = 0;
1306
1307     /* Start outer loop over neighborlists */
1308     for(iidx=0; iidx<nri; iidx++)
1309     {
1310         /* Load shift vector for this list */
1311         i_shift_offset   = DIM*shiftidx[iidx];
1312
1313         /* Load limits for loop over neighbors */
1314         j_index_start    = jindex[iidx];
1315         j_index_end      = jindex[iidx+1];
1316
1317         /* Get outer coordinate index */
1318         inr              = iinr[iidx];
1319         i_coord_offset   = DIM*inr;
1320
1321         /* Load i particle coords and add shift vector */
1322         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1323                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1324
1325         fix0             = _mm_setzero_pd();
1326         fiy0             = _mm_setzero_pd();
1327         fiz0             = _mm_setzero_pd();
1328         fix1             = _mm_setzero_pd();
1329         fiy1             = _mm_setzero_pd();
1330         fiz1             = _mm_setzero_pd();
1331         fix2             = _mm_setzero_pd();
1332         fiy2             = _mm_setzero_pd();
1333         fiz2             = _mm_setzero_pd();
1334
1335         /* Start inner kernel loop */
1336         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1337         {
1338
1339             /* Get j neighbor index, and coordinate index */
1340             jnrA             = jjnr[jidx];
1341             jnrB             = jjnr[jidx+1];
1342             j_coord_offsetA  = DIM*jnrA;
1343             j_coord_offsetB  = DIM*jnrB;
1344
1345             /* load j atom coordinates */
1346             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1347                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1348
1349             /* Calculate displacement vector */
1350             dx00             = _mm_sub_pd(ix0,jx0);
1351             dy00             = _mm_sub_pd(iy0,jy0);
1352             dz00             = _mm_sub_pd(iz0,jz0);
1353             dx01             = _mm_sub_pd(ix0,jx1);
1354             dy01             = _mm_sub_pd(iy0,jy1);
1355             dz01             = _mm_sub_pd(iz0,jz1);
1356             dx02             = _mm_sub_pd(ix0,jx2);
1357             dy02             = _mm_sub_pd(iy0,jy2);
1358             dz02             = _mm_sub_pd(iz0,jz2);
1359             dx10             = _mm_sub_pd(ix1,jx0);
1360             dy10             = _mm_sub_pd(iy1,jy0);
1361             dz10             = _mm_sub_pd(iz1,jz0);
1362             dx11             = _mm_sub_pd(ix1,jx1);
1363             dy11             = _mm_sub_pd(iy1,jy1);
1364             dz11             = _mm_sub_pd(iz1,jz1);
1365             dx12             = _mm_sub_pd(ix1,jx2);
1366             dy12             = _mm_sub_pd(iy1,jy2);
1367             dz12             = _mm_sub_pd(iz1,jz2);
1368             dx20             = _mm_sub_pd(ix2,jx0);
1369             dy20             = _mm_sub_pd(iy2,jy0);
1370             dz20             = _mm_sub_pd(iz2,jz0);
1371             dx21             = _mm_sub_pd(ix2,jx1);
1372             dy21             = _mm_sub_pd(iy2,jy1);
1373             dz21             = _mm_sub_pd(iz2,jz1);
1374             dx22             = _mm_sub_pd(ix2,jx2);
1375             dy22             = _mm_sub_pd(iy2,jy2);
1376             dz22             = _mm_sub_pd(iz2,jz2);
1377
1378             /* Calculate squared distance and things based on it */
1379             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1380             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1381             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1382             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1383             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1384             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1385             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1386             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1387             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1388
1389             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1390             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1391             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1392             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1393             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1394             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1395             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1396             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1397             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1398
1399             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1400             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1401             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1402             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1403             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1404             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1405             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1406             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1407             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1408
1409             fjx0             = _mm_setzero_pd();
1410             fjy0             = _mm_setzero_pd();
1411             fjz0             = _mm_setzero_pd();
1412             fjx1             = _mm_setzero_pd();
1413             fjy1             = _mm_setzero_pd();
1414             fjz1             = _mm_setzero_pd();
1415             fjx2             = _mm_setzero_pd();
1416             fjy2             = _mm_setzero_pd();
1417             fjz2             = _mm_setzero_pd();
1418
1419             /**************************
1420              * CALCULATE INTERACTIONS *
1421              **************************/
1422
1423             if (gmx_mm_any_lt(rsq00,rcutoff2))
1424             {
1425
1426             r00              = _mm_mul_pd(rsq00,rinv00);
1427
1428             /* Calculate table index by multiplying r with table scale and truncate to integer */
1429             rt               = _mm_mul_pd(r00,vftabscale);
1430             vfitab           = _mm_cvttpd_epi32(rt);
1431             vfeps            = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
1432             vfitab           = _mm_slli_epi32(vfitab,3);
1433
1434             /* REACTION-FIELD ELECTROSTATICS */
1435             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
1436
1437             /* CUBIC SPLINE TABLE DISPERSION */
1438             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1439             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
1440             GMX_MM_TRANSPOSE2_PD(Y,F);
1441             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1442             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
1443             GMX_MM_TRANSPOSE2_PD(G,H);
1444             Heps             = _mm_mul_pd(vfeps,H);
1445             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1446             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1447             fvdw6            = _mm_mul_pd(c6_00,FF);
1448
1449             /* CUBIC SPLINE TABLE REPULSION */
1450             vfitab           = _mm_add_epi32(vfitab,ifour);
1451             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1452             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
1453             GMX_MM_TRANSPOSE2_PD(Y,F);
1454             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1455             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
1456             GMX_MM_TRANSPOSE2_PD(G,H);
1457             Heps             = _mm_mul_pd(vfeps,H);
1458             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1459             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1460             fvdw12           = _mm_mul_pd(c12_00,FF);
1461             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1462
1463             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1464
1465             fscal            = _mm_add_pd(felec,fvdw);
1466
1467             fscal            = _mm_and_pd(fscal,cutoff_mask);
1468
1469             /* Calculate temporary vectorial force */
1470             tx               = _mm_mul_pd(fscal,dx00);
1471             ty               = _mm_mul_pd(fscal,dy00);
1472             tz               = _mm_mul_pd(fscal,dz00);
1473
1474             /* Update vectorial force */
1475             fix0             = _mm_add_pd(fix0,tx);
1476             fiy0             = _mm_add_pd(fiy0,ty);
1477             fiz0             = _mm_add_pd(fiz0,tz);
1478
1479             fjx0             = _mm_add_pd(fjx0,tx);
1480             fjy0             = _mm_add_pd(fjy0,ty);
1481             fjz0             = _mm_add_pd(fjz0,tz);
1482
1483             }
1484
1485             /**************************
1486              * CALCULATE INTERACTIONS *
1487              **************************/
1488
1489             if (gmx_mm_any_lt(rsq01,rcutoff2))
1490             {
1491
1492             /* REACTION-FIELD ELECTROSTATICS */
1493             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
1494
1495             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
1496
1497             fscal            = felec;
1498
1499             fscal            = _mm_and_pd(fscal,cutoff_mask);
1500
1501             /* Calculate temporary vectorial force */
1502             tx               = _mm_mul_pd(fscal,dx01);
1503             ty               = _mm_mul_pd(fscal,dy01);
1504             tz               = _mm_mul_pd(fscal,dz01);
1505
1506             /* Update vectorial force */
1507             fix0             = _mm_add_pd(fix0,tx);
1508             fiy0             = _mm_add_pd(fiy0,ty);
1509             fiz0             = _mm_add_pd(fiz0,tz);
1510
1511             fjx1             = _mm_add_pd(fjx1,tx);
1512             fjy1             = _mm_add_pd(fjy1,ty);
1513             fjz1             = _mm_add_pd(fjz1,tz);
1514
1515             }
1516
1517             /**************************
1518              * CALCULATE INTERACTIONS *
1519              **************************/
1520
1521             if (gmx_mm_any_lt(rsq02,rcutoff2))
1522             {
1523
1524             /* REACTION-FIELD ELECTROSTATICS */
1525             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
1526
1527             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
1528
1529             fscal            = felec;
1530
1531             fscal            = _mm_and_pd(fscal,cutoff_mask);
1532
1533             /* Calculate temporary vectorial force */
1534             tx               = _mm_mul_pd(fscal,dx02);
1535             ty               = _mm_mul_pd(fscal,dy02);
1536             tz               = _mm_mul_pd(fscal,dz02);
1537
1538             /* Update vectorial force */
1539             fix0             = _mm_add_pd(fix0,tx);
1540             fiy0             = _mm_add_pd(fiy0,ty);
1541             fiz0             = _mm_add_pd(fiz0,tz);
1542
1543             fjx2             = _mm_add_pd(fjx2,tx);
1544             fjy2             = _mm_add_pd(fjy2,ty);
1545             fjz2             = _mm_add_pd(fjz2,tz);
1546
1547             }
1548
1549             /**************************
1550              * CALCULATE INTERACTIONS *
1551              **************************/
1552
1553             if (gmx_mm_any_lt(rsq10,rcutoff2))
1554             {
1555
1556             /* REACTION-FIELD ELECTROSTATICS */
1557             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
1558
1559             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
1560
1561             fscal            = felec;
1562
1563             fscal            = _mm_and_pd(fscal,cutoff_mask);
1564
1565             /* Calculate temporary vectorial force */
1566             tx               = _mm_mul_pd(fscal,dx10);
1567             ty               = _mm_mul_pd(fscal,dy10);
1568             tz               = _mm_mul_pd(fscal,dz10);
1569
1570             /* Update vectorial force */
1571             fix1             = _mm_add_pd(fix1,tx);
1572             fiy1             = _mm_add_pd(fiy1,ty);
1573             fiz1             = _mm_add_pd(fiz1,tz);
1574
1575             fjx0             = _mm_add_pd(fjx0,tx);
1576             fjy0             = _mm_add_pd(fjy0,ty);
1577             fjz0             = _mm_add_pd(fjz0,tz);
1578
1579             }
1580
1581             /**************************
1582              * CALCULATE INTERACTIONS *
1583              **************************/
1584
1585             if (gmx_mm_any_lt(rsq11,rcutoff2))
1586             {
1587
1588             /* REACTION-FIELD ELECTROSTATICS */
1589             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1590
1591             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1592
1593             fscal            = felec;
1594
1595             fscal            = _mm_and_pd(fscal,cutoff_mask);
1596
1597             /* Calculate temporary vectorial force */
1598             tx               = _mm_mul_pd(fscal,dx11);
1599             ty               = _mm_mul_pd(fscal,dy11);
1600             tz               = _mm_mul_pd(fscal,dz11);
1601
1602             /* Update vectorial force */
1603             fix1             = _mm_add_pd(fix1,tx);
1604             fiy1             = _mm_add_pd(fiy1,ty);
1605             fiz1             = _mm_add_pd(fiz1,tz);
1606
1607             fjx1             = _mm_add_pd(fjx1,tx);
1608             fjy1             = _mm_add_pd(fjy1,ty);
1609             fjz1             = _mm_add_pd(fjz1,tz);
1610
1611             }
1612
1613             /**************************
1614              * CALCULATE INTERACTIONS *
1615              **************************/
1616
1617             if (gmx_mm_any_lt(rsq12,rcutoff2))
1618             {
1619
1620             /* REACTION-FIELD ELECTROSTATICS */
1621             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1622
1623             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1624
1625             fscal            = felec;
1626
1627             fscal            = _mm_and_pd(fscal,cutoff_mask);
1628
1629             /* Calculate temporary vectorial force */
1630             tx               = _mm_mul_pd(fscal,dx12);
1631             ty               = _mm_mul_pd(fscal,dy12);
1632             tz               = _mm_mul_pd(fscal,dz12);
1633
1634             /* Update vectorial force */
1635             fix1             = _mm_add_pd(fix1,tx);
1636             fiy1             = _mm_add_pd(fiy1,ty);
1637             fiz1             = _mm_add_pd(fiz1,tz);
1638
1639             fjx2             = _mm_add_pd(fjx2,tx);
1640             fjy2             = _mm_add_pd(fjy2,ty);
1641             fjz2             = _mm_add_pd(fjz2,tz);
1642
1643             }
1644
1645             /**************************
1646              * CALCULATE INTERACTIONS *
1647              **************************/
1648
1649             if (gmx_mm_any_lt(rsq20,rcutoff2))
1650             {
1651
1652             /* REACTION-FIELD ELECTROSTATICS */
1653             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
1654
1655             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
1656
1657             fscal            = felec;
1658
1659             fscal            = _mm_and_pd(fscal,cutoff_mask);
1660
1661             /* Calculate temporary vectorial force */
1662             tx               = _mm_mul_pd(fscal,dx20);
1663             ty               = _mm_mul_pd(fscal,dy20);
1664             tz               = _mm_mul_pd(fscal,dz20);
1665
1666             /* Update vectorial force */
1667             fix2             = _mm_add_pd(fix2,tx);
1668             fiy2             = _mm_add_pd(fiy2,ty);
1669             fiz2             = _mm_add_pd(fiz2,tz);
1670
1671             fjx0             = _mm_add_pd(fjx0,tx);
1672             fjy0             = _mm_add_pd(fjy0,ty);
1673             fjz0             = _mm_add_pd(fjz0,tz);
1674
1675             }
1676
1677             /**************************
1678              * CALCULATE INTERACTIONS *
1679              **************************/
1680
1681             if (gmx_mm_any_lt(rsq21,rcutoff2))
1682             {
1683
1684             /* REACTION-FIELD ELECTROSTATICS */
1685             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1686
1687             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1688
1689             fscal            = felec;
1690
1691             fscal            = _mm_and_pd(fscal,cutoff_mask);
1692
1693             /* Calculate temporary vectorial force */
1694             tx               = _mm_mul_pd(fscal,dx21);
1695             ty               = _mm_mul_pd(fscal,dy21);
1696             tz               = _mm_mul_pd(fscal,dz21);
1697
1698             /* Update vectorial force */
1699             fix2             = _mm_add_pd(fix2,tx);
1700             fiy2             = _mm_add_pd(fiy2,ty);
1701             fiz2             = _mm_add_pd(fiz2,tz);
1702
1703             fjx1             = _mm_add_pd(fjx1,tx);
1704             fjy1             = _mm_add_pd(fjy1,ty);
1705             fjz1             = _mm_add_pd(fjz1,tz);
1706
1707             }
1708
1709             /**************************
1710              * CALCULATE INTERACTIONS *
1711              **************************/
1712
1713             if (gmx_mm_any_lt(rsq22,rcutoff2))
1714             {
1715
1716             /* REACTION-FIELD ELECTROSTATICS */
1717             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1718
1719             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1720
1721             fscal            = felec;
1722
1723             fscal            = _mm_and_pd(fscal,cutoff_mask);
1724
1725             /* Calculate temporary vectorial force */
1726             tx               = _mm_mul_pd(fscal,dx22);
1727             ty               = _mm_mul_pd(fscal,dy22);
1728             tz               = _mm_mul_pd(fscal,dz22);
1729
1730             /* Update vectorial force */
1731             fix2             = _mm_add_pd(fix2,tx);
1732             fiy2             = _mm_add_pd(fiy2,ty);
1733             fiz2             = _mm_add_pd(fiz2,tz);
1734
1735             fjx2             = _mm_add_pd(fjx2,tx);
1736             fjy2             = _mm_add_pd(fjy2,ty);
1737             fjz2             = _mm_add_pd(fjz2,tz);
1738
1739             }
1740
1741             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1742
1743             /* Inner loop uses 297 flops */
1744         }
1745
1746         if(jidx<j_index_end)
1747         {
1748
1749             jnrA             = jjnr[jidx];
1750             j_coord_offsetA  = DIM*jnrA;
1751
1752             /* load j atom coordinates */
1753             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1754                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1755
1756             /* Calculate displacement vector */
1757             dx00             = _mm_sub_pd(ix0,jx0);
1758             dy00             = _mm_sub_pd(iy0,jy0);
1759             dz00             = _mm_sub_pd(iz0,jz0);
1760             dx01             = _mm_sub_pd(ix0,jx1);
1761             dy01             = _mm_sub_pd(iy0,jy1);
1762             dz01             = _mm_sub_pd(iz0,jz1);
1763             dx02             = _mm_sub_pd(ix0,jx2);
1764             dy02             = _mm_sub_pd(iy0,jy2);
1765             dz02             = _mm_sub_pd(iz0,jz2);
1766             dx10             = _mm_sub_pd(ix1,jx0);
1767             dy10             = _mm_sub_pd(iy1,jy0);
1768             dz10             = _mm_sub_pd(iz1,jz0);
1769             dx11             = _mm_sub_pd(ix1,jx1);
1770             dy11             = _mm_sub_pd(iy1,jy1);
1771             dz11             = _mm_sub_pd(iz1,jz1);
1772             dx12             = _mm_sub_pd(ix1,jx2);
1773             dy12             = _mm_sub_pd(iy1,jy2);
1774             dz12             = _mm_sub_pd(iz1,jz2);
1775             dx20             = _mm_sub_pd(ix2,jx0);
1776             dy20             = _mm_sub_pd(iy2,jy0);
1777             dz20             = _mm_sub_pd(iz2,jz0);
1778             dx21             = _mm_sub_pd(ix2,jx1);
1779             dy21             = _mm_sub_pd(iy2,jy1);
1780             dz21             = _mm_sub_pd(iz2,jz1);
1781             dx22             = _mm_sub_pd(ix2,jx2);
1782             dy22             = _mm_sub_pd(iy2,jy2);
1783             dz22             = _mm_sub_pd(iz2,jz2);
1784
1785             /* Calculate squared distance and things based on it */
1786             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1787             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1788             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1789             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1790             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1791             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1792             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1793             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1794             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1795
1796             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1797             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1798             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1799             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1800             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1801             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1802             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1803             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1804             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1805
1806             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1807             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1808             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1809             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1810             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1811             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1812             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1813             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1814             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1815
1816             fjx0             = _mm_setzero_pd();
1817             fjy0             = _mm_setzero_pd();
1818             fjz0             = _mm_setzero_pd();
1819             fjx1             = _mm_setzero_pd();
1820             fjy1             = _mm_setzero_pd();
1821             fjz1             = _mm_setzero_pd();
1822             fjx2             = _mm_setzero_pd();
1823             fjy2             = _mm_setzero_pd();
1824             fjz2             = _mm_setzero_pd();
1825
1826             /**************************
1827              * CALCULATE INTERACTIONS *
1828              **************************/
1829
1830             if (gmx_mm_any_lt(rsq00,rcutoff2))
1831             {
1832
1833             r00              = _mm_mul_pd(rsq00,rinv00);
1834
1835             /* Calculate table index by multiplying r with table scale and truncate to integer */
1836             rt               = _mm_mul_pd(r00,vftabscale);
1837             vfitab           = _mm_cvttpd_epi32(rt);
1838             vfeps            = _mm_sub_pd(rt,_mm_cvtepi32_pd(vfitab));
1839             vfitab           = _mm_slli_epi32(vfitab,3);
1840
1841             /* REACTION-FIELD ELECTROSTATICS */
1842             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
1843
1844             /* CUBIC SPLINE TABLE DISPERSION */
1845             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1846             F                = _mm_setzero_pd();
1847             GMX_MM_TRANSPOSE2_PD(Y,F);
1848             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1849             H                = _mm_setzero_pd();
1850             GMX_MM_TRANSPOSE2_PD(G,H);
1851             Heps             = _mm_mul_pd(vfeps,H);
1852             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1853             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1854             fvdw6            = _mm_mul_pd(c6_00,FF);
1855
1856             /* CUBIC SPLINE TABLE REPULSION */
1857             vfitab           = _mm_add_epi32(vfitab,ifour);
1858             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1859             F                = _mm_setzero_pd();
1860             GMX_MM_TRANSPOSE2_PD(Y,F);
1861             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1862             H                = _mm_setzero_pd();
1863             GMX_MM_TRANSPOSE2_PD(G,H);
1864             Heps             = _mm_mul_pd(vfeps,H);
1865             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1866             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1867             fvdw12           = _mm_mul_pd(c12_00,FF);
1868             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1869
1870             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1871
1872             fscal            = _mm_add_pd(felec,fvdw);
1873
1874             fscal            = _mm_and_pd(fscal,cutoff_mask);
1875
1876             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1877
1878             /* Calculate temporary vectorial force */
1879             tx               = _mm_mul_pd(fscal,dx00);
1880             ty               = _mm_mul_pd(fscal,dy00);
1881             tz               = _mm_mul_pd(fscal,dz00);
1882
1883             /* Update vectorial force */
1884             fix0             = _mm_add_pd(fix0,tx);
1885             fiy0             = _mm_add_pd(fiy0,ty);
1886             fiz0             = _mm_add_pd(fiz0,tz);
1887
1888             fjx0             = _mm_add_pd(fjx0,tx);
1889             fjy0             = _mm_add_pd(fjy0,ty);
1890             fjz0             = _mm_add_pd(fjz0,tz);
1891
1892             }
1893
1894             /**************************
1895              * CALCULATE INTERACTIONS *
1896              **************************/
1897
1898             if (gmx_mm_any_lt(rsq01,rcutoff2))
1899             {
1900
1901             /* REACTION-FIELD ELECTROSTATICS */
1902             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
1903
1904             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
1905
1906             fscal            = felec;
1907
1908             fscal            = _mm_and_pd(fscal,cutoff_mask);
1909
1910             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1911
1912             /* Calculate temporary vectorial force */
1913             tx               = _mm_mul_pd(fscal,dx01);
1914             ty               = _mm_mul_pd(fscal,dy01);
1915             tz               = _mm_mul_pd(fscal,dz01);
1916
1917             /* Update vectorial force */
1918             fix0             = _mm_add_pd(fix0,tx);
1919             fiy0             = _mm_add_pd(fiy0,ty);
1920             fiz0             = _mm_add_pd(fiz0,tz);
1921
1922             fjx1             = _mm_add_pd(fjx1,tx);
1923             fjy1             = _mm_add_pd(fjy1,ty);
1924             fjz1             = _mm_add_pd(fjz1,tz);
1925
1926             }
1927
1928             /**************************
1929              * CALCULATE INTERACTIONS *
1930              **************************/
1931
1932             if (gmx_mm_any_lt(rsq02,rcutoff2))
1933             {
1934
1935             /* REACTION-FIELD ELECTROSTATICS */
1936             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
1937
1938             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
1939
1940             fscal            = felec;
1941
1942             fscal            = _mm_and_pd(fscal,cutoff_mask);
1943
1944             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1945
1946             /* Calculate temporary vectorial force */
1947             tx               = _mm_mul_pd(fscal,dx02);
1948             ty               = _mm_mul_pd(fscal,dy02);
1949             tz               = _mm_mul_pd(fscal,dz02);
1950
1951             /* Update vectorial force */
1952             fix0             = _mm_add_pd(fix0,tx);
1953             fiy0             = _mm_add_pd(fiy0,ty);
1954             fiz0             = _mm_add_pd(fiz0,tz);
1955
1956             fjx2             = _mm_add_pd(fjx2,tx);
1957             fjy2             = _mm_add_pd(fjy2,ty);
1958             fjz2             = _mm_add_pd(fjz2,tz);
1959
1960             }
1961
1962             /**************************
1963              * CALCULATE INTERACTIONS *
1964              **************************/
1965
1966             if (gmx_mm_any_lt(rsq10,rcutoff2))
1967             {
1968
1969             /* REACTION-FIELD ELECTROSTATICS */
1970             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
1971
1972             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
1973
1974             fscal            = felec;
1975
1976             fscal            = _mm_and_pd(fscal,cutoff_mask);
1977
1978             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1979
1980             /* Calculate temporary vectorial force */
1981             tx               = _mm_mul_pd(fscal,dx10);
1982             ty               = _mm_mul_pd(fscal,dy10);
1983             tz               = _mm_mul_pd(fscal,dz10);
1984
1985             /* Update vectorial force */
1986             fix1             = _mm_add_pd(fix1,tx);
1987             fiy1             = _mm_add_pd(fiy1,ty);
1988             fiz1             = _mm_add_pd(fiz1,tz);
1989
1990             fjx0             = _mm_add_pd(fjx0,tx);
1991             fjy0             = _mm_add_pd(fjy0,ty);
1992             fjz0             = _mm_add_pd(fjz0,tz);
1993
1994             }
1995
1996             /**************************
1997              * CALCULATE INTERACTIONS *
1998              **************************/
1999
2000             if (gmx_mm_any_lt(rsq11,rcutoff2))
2001             {
2002
2003             /* REACTION-FIELD ELECTROSTATICS */
2004             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
2005
2006             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
2007
2008             fscal            = felec;
2009
2010             fscal            = _mm_and_pd(fscal,cutoff_mask);
2011
2012             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2013
2014             /* Calculate temporary vectorial force */
2015             tx               = _mm_mul_pd(fscal,dx11);
2016             ty               = _mm_mul_pd(fscal,dy11);
2017             tz               = _mm_mul_pd(fscal,dz11);
2018
2019             /* Update vectorial force */
2020             fix1             = _mm_add_pd(fix1,tx);
2021             fiy1             = _mm_add_pd(fiy1,ty);
2022             fiz1             = _mm_add_pd(fiz1,tz);
2023
2024             fjx1             = _mm_add_pd(fjx1,tx);
2025             fjy1             = _mm_add_pd(fjy1,ty);
2026             fjz1             = _mm_add_pd(fjz1,tz);
2027
2028             }
2029
2030             /**************************
2031              * CALCULATE INTERACTIONS *
2032              **************************/
2033
2034             if (gmx_mm_any_lt(rsq12,rcutoff2))
2035             {
2036
2037             /* REACTION-FIELD ELECTROSTATICS */
2038             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
2039
2040             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
2041
2042             fscal            = felec;
2043
2044             fscal            = _mm_and_pd(fscal,cutoff_mask);
2045
2046             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2047
2048             /* Calculate temporary vectorial force */
2049             tx               = _mm_mul_pd(fscal,dx12);
2050             ty               = _mm_mul_pd(fscal,dy12);
2051             tz               = _mm_mul_pd(fscal,dz12);
2052
2053             /* Update vectorial force */
2054             fix1             = _mm_add_pd(fix1,tx);
2055             fiy1             = _mm_add_pd(fiy1,ty);
2056             fiz1             = _mm_add_pd(fiz1,tz);
2057
2058             fjx2             = _mm_add_pd(fjx2,tx);
2059             fjy2             = _mm_add_pd(fjy2,ty);
2060             fjz2             = _mm_add_pd(fjz2,tz);
2061
2062             }
2063
2064             /**************************
2065              * CALCULATE INTERACTIONS *
2066              **************************/
2067
2068             if (gmx_mm_any_lt(rsq20,rcutoff2))
2069             {
2070
2071             /* REACTION-FIELD ELECTROSTATICS */
2072             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
2073
2074             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
2075
2076             fscal            = felec;
2077
2078             fscal            = _mm_and_pd(fscal,cutoff_mask);
2079
2080             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2081
2082             /* Calculate temporary vectorial force */
2083             tx               = _mm_mul_pd(fscal,dx20);
2084             ty               = _mm_mul_pd(fscal,dy20);
2085             tz               = _mm_mul_pd(fscal,dz20);
2086
2087             /* Update vectorial force */
2088             fix2             = _mm_add_pd(fix2,tx);
2089             fiy2             = _mm_add_pd(fiy2,ty);
2090             fiz2             = _mm_add_pd(fiz2,tz);
2091
2092             fjx0             = _mm_add_pd(fjx0,tx);
2093             fjy0             = _mm_add_pd(fjy0,ty);
2094             fjz0             = _mm_add_pd(fjz0,tz);
2095
2096             }
2097
2098             /**************************
2099              * CALCULATE INTERACTIONS *
2100              **************************/
2101
2102             if (gmx_mm_any_lt(rsq21,rcutoff2))
2103             {
2104
2105             /* REACTION-FIELD ELECTROSTATICS */
2106             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
2107
2108             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
2109
2110             fscal            = felec;
2111
2112             fscal            = _mm_and_pd(fscal,cutoff_mask);
2113
2114             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2115
2116             /* Calculate temporary vectorial force */
2117             tx               = _mm_mul_pd(fscal,dx21);
2118             ty               = _mm_mul_pd(fscal,dy21);
2119             tz               = _mm_mul_pd(fscal,dz21);
2120
2121             /* Update vectorial force */
2122             fix2             = _mm_add_pd(fix2,tx);
2123             fiy2             = _mm_add_pd(fiy2,ty);
2124             fiz2             = _mm_add_pd(fiz2,tz);
2125
2126             fjx1             = _mm_add_pd(fjx1,tx);
2127             fjy1             = _mm_add_pd(fjy1,ty);
2128             fjz1             = _mm_add_pd(fjz1,tz);
2129
2130             }
2131
2132             /**************************
2133              * CALCULATE INTERACTIONS *
2134              **************************/
2135
2136             if (gmx_mm_any_lt(rsq22,rcutoff2))
2137             {
2138
2139             /* REACTION-FIELD ELECTROSTATICS */
2140             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
2141
2142             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
2143
2144             fscal            = felec;
2145
2146             fscal            = _mm_and_pd(fscal,cutoff_mask);
2147
2148             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2149
2150             /* Calculate temporary vectorial force */
2151             tx               = _mm_mul_pd(fscal,dx22);
2152             ty               = _mm_mul_pd(fscal,dy22);
2153             tz               = _mm_mul_pd(fscal,dz22);
2154
2155             /* Update vectorial force */
2156             fix2             = _mm_add_pd(fix2,tx);
2157             fiy2             = _mm_add_pd(fiy2,ty);
2158             fiz2             = _mm_add_pd(fiz2,tz);
2159
2160             fjx2             = _mm_add_pd(fjx2,tx);
2161             fjy2             = _mm_add_pd(fjy2,ty);
2162             fjz2             = _mm_add_pd(fjz2,tz);
2163
2164             }
2165
2166             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2167
2168             /* Inner loop uses 297 flops */
2169         }
2170
2171         /* End of innermost loop */
2172
2173         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
2174                                               f+i_coord_offset,fshift+i_shift_offset);
2175
2176         /* Increment number of inner iterations */
2177         inneriter                  += j_index_end - j_index_start;
2178
2179         /* Outer loop uses 18 flops */
2180     }
2181
2182     /* Increment number of outer iterations */
2183     outeriter        += nri;
2184
2185     /* Update outer/inner flops */
2186
2187     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*297);
2188 }