src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecRFCut_VdwCSTab_GeomW4W4_avx_128_fma_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_double kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_double.h"
  48 #include "kernelutil_x86_avx_128_fma_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW4W4_VF_avx_128_fma_double
  52  * Electrostatics interaction: ReactionField
  53  * VdW interaction:            CubicSplineTable
  54  * Geometry:                   Water4-Water4
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecRFCut_VdwCSTab_GeomW4W4_VF_avx_128_fma_double
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB;
  75     int              j_coord_offsetA,j_coord_offsetB;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  80     int              vdwioffset0;
  81     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  82     int              vdwioffset1;
  83     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  84     int              vdwioffset2;
  85     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  86     int              vdwioffset3;
  87     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  88     int              vdwjidx0A,vdwjidx0B;
  89     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     int              vdwjidx1A,vdwjidx1B;
  91     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  92     int              vdwjidx2A,vdwjidx2B;
  93     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  94     int              vdwjidx3A,vdwjidx3B;
  95     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  96     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  97     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  98     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  99     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 100     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 101     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 102     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 103     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 104     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 105     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 106     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 107     real             *charge;
 108     int              nvdwtype;
 109     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 110     int              *vdwtype;
 111     real             *vdwparam;
 112     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 113     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 114     __m128i          vfitab;
 115     __m128i          ifour       = _mm_set1_epi32(4);
 116     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
 117     real             *vftab;
 118     __m128d          dummy_mask,cutoff_mask;
 119     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 120     __m128d          one     = _mm_set1_pd(1.0);
 121     __m128d          two     = _mm_set1_pd(2.0);
 122     x                = xx[0];
 123     f                = ff[0];
 124
 125     nri              = nlist->nri;
 126     iinr             = nlist->iinr;
 127     jindex           = nlist->jindex;
 128     jjnr             = nlist->jjnr;
 129     shiftidx         = nlist->shift;
 130     gid              = nlist->gid;
 131     shiftvec         = fr->shift_vec[0];
 132     fshift           = fr->fshift[0];
 133     facel            = _mm_set1_pd(fr->epsfac);
 134     charge           = mdatoms->chargeA;
 135     krf              = _mm_set1_pd(fr->ic->k_rf);
 136     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 137     crf              = _mm_set1_pd(fr->ic->c_rf);
 138     nvdwtype         = fr->ntype;
 139     vdwparam         = fr->nbfp;
 140     vdwtype          = mdatoms->typeA;
 141
 142     vftab            = kernel_data->table_vdw->data;
 143     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
 144
 145     /* Setup water-specific parameters */
 146     inr              = nlist->iinr[0];
 147     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 148     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 149     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 150     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 151
 152     jq1              = _mm_set1_pd(charge[inr+1]);
 153     jq2              = _mm_set1_pd(charge[inr+2]);
 154     jq3              = _mm_set1_pd(charge[inr+3]);
 155     vdwjidx0A        = 2*vdwtype[inr+0];
 156     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 157     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 158     qq11             = _mm_mul_pd(iq1,jq1);
 159     qq12             = _mm_mul_pd(iq1,jq2);
 160     qq13             = _mm_mul_pd(iq1,jq3);
 161     qq21             = _mm_mul_pd(iq2,jq1);
 162     qq22             = _mm_mul_pd(iq2,jq2);
 163     qq23             = _mm_mul_pd(iq2,jq3);
 164     qq31             = _mm_mul_pd(iq3,jq1);
 165     qq32             = _mm_mul_pd(iq3,jq2);
 166     qq33             = _mm_mul_pd(iq3,jq3);
 167
 168     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 169     rcutoff_scalar   = fr->rcoulomb;
 170     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 171     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 172
 173     /* Avoid stupid compiler warnings */
 174     jnrA = jnrB = 0;
 175     j_coord_offsetA = 0;
 176     j_coord_offsetB = 0;
 177
 178     outeriter        = 0;
 179     inneriter        = 0;
 180
 181     /* Start outer loop over neighborlists */
 182     for(iidx=0; iidx<nri; iidx++)
 183     {
 184         /* Load shift vector for this list */
 185         i_shift_offset   = DIM*shiftidx[iidx];
 186
 187         /* Load limits for loop over neighbors */
 188         j_index_start    = jindex[iidx];
 189         j_index_end      = jindex[iidx+1];
 190
 191         /* Get outer coordinate index */
 192         inr              = iinr[iidx];
 193         i_coord_offset   = DIM*inr;
 194
 195         /* Load i particle coords and add shift vector */
 196         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 197                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 198
 199         fix0             = _mm_setzero_pd();
 200         fiy0             = _mm_setzero_pd();
 201         fiz0             = _mm_setzero_pd();
 202         fix1             = _mm_setzero_pd();
 203         fiy1             = _mm_setzero_pd();
 204         fiz1             = _mm_setzero_pd();
 205         fix2             = _mm_setzero_pd();
 206         fiy2             = _mm_setzero_pd();
 207         fiz2             = _mm_setzero_pd();
 208         fix3             = _mm_setzero_pd();
 209         fiy3             = _mm_setzero_pd();
 210         fiz3             = _mm_setzero_pd();
 211
 212         /* Reset potential sums */
 213         velecsum         = _mm_setzero_pd();
 214         vvdwsum          = _mm_setzero_pd();
 215
 216         /* Start inner kernel loop */
 217         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 218         {
 219
 220             /* Get j neighbor index, and coordinate index */
 221             jnrA             = jjnr[jidx];
 222             jnrB             = jjnr[jidx+1];
 223             j_coord_offsetA  = DIM*jnrA;
 224             j_coord_offsetB  = DIM*jnrB;
 225
 226             /* load j atom coordinates */
 227             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 228                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 229                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 230
 231             /* Calculate displacement vector */
 232             dx00             = _mm_sub_pd(ix0,jx0);
 233             dy00             = _mm_sub_pd(iy0,jy0);
 234             dz00             = _mm_sub_pd(iz0,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             dx13             = _mm_sub_pd(ix1,jx3);
 242             dy13             = _mm_sub_pd(iy1,jy3);
 243             dz13             = _mm_sub_pd(iz1,jz3);
 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             dx23             = _mm_sub_pd(ix2,jx3);
 251             dy23             = _mm_sub_pd(iy2,jy3);
 252             dz23             = _mm_sub_pd(iz2,jz3);
 253             dx31             = _mm_sub_pd(ix3,jx1);
 254             dy31             = _mm_sub_pd(iy3,jy1);
 255             dz31             = _mm_sub_pd(iz3,jz1);
 256             dx32             = _mm_sub_pd(ix3,jx2);
 257             dy32             = _mm_sub_pd(iy3,jy2);
 258             dz32             = _mm_sub_pd(iz3,jz2);
 259             dx33             = _mm_sub_pd(ix3,jx3);
 260             dy33             = _mm_sub_pd(iy3,jy3);
 261             dz33             = _mm_sub_pd(iz3,jz3);
 262
 263             /* Calculate squared distance and things based on it */
 264             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 265             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 266             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 267             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 268             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 269             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 270             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 271             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 272             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 273             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 274
 275             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 276             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 277             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 278             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 279             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 280             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 281             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 282             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 283             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 284             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 285
 286             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 287             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 288             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 289             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 290             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 291             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 292             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 293             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 294             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 295
 296             fjx0             = _mm_setzero_pd();
 297             fjy0             = _mm_setzero_pd();
 298             fjz0             = _mm_setzero_pd();
 299             fjx1             = _mm_setzero_pd();
 300             fjy1             = _mm_setzero_pd();
 301             fjz1             = _mm_setzero_pd();
 302             fjx2             = _mm_setzero_pd();
 303             fjy2             = _mm_setzero_pd();
 304             fjz2             = _mm_setzero_pd();
 305             fjx3             = _mm_setzero_pd();
 306             fjy3             = _mm_setzero_pd();
 307             fjz3             = _mm_setzero_pd();
 308
 309             /**************************
 310              * CALCULATE INTERACTIONS *
 311              **************************/
 312
 313             r00              = _mm_mul_pd(rsq00,rinv00);
 314
 315             /* Calculate table index by multiplying r with table scale and truncate to integer */
 316             rt               = _mm_mul_pd(r00,vftabscale);
 317             vfitab           = _mm_cvttpd_epi32(rt);
 318 #ifdef __XOP__
 319             vfeps            = _mm_frcz_pd(rt);
 320 #else
 321             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 322 #endif
 323             twovfeps         = _mm_add_pd(vfeps,vfeps);
 324             vfitab           = _mm_slli_epi32(vfitab,3);
 325
 326             /* CUBIC SPLINE TABLE DISPERSION */
 327             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 328             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 329             GMX_MM_TRANSPOSE2_PD(Y,F);
 330             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 331             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
 332             GMX_MM_TRANSPOSE2_PD(G,H);
 333             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 334             VV               = _mm_macc_pd(vfeps,Fp,Y);
 335             vvdw6            = _mm_mul_pd(c6_00,VV);
 336             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 337             fvdw6            = _mm_mul_pd(c6_00,FF);
 338
 339             /* CUBIC SPLINE TABLE REPULSION */
 340             vfitab           = _mm_add_epi32(vfitab,ifour);
 341             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 342             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 343             GMX_MM_TRANSPOSE2_PD(Y,F);
 344             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 345             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
 346             GMX_MM_TRANSPOSE2_PD(G,H);
 347             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 348             VV               = _mm_macc_pd(vfeps,Fp,Y);
 349             vvdw12           = _mm_mul_pd(c12_00,VV);
 350             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 351             fvdw12           = _mm_mul_pd(c12_00,FF);
 352             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 353             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 354
 355             /* Update potential sum for this i atom from the interaction with this j atom. */
 356             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 357
 358             fscal            = fvdw;
 359
 360             /* Update vectorial force */
 361             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 362             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 363             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 364
 365             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 366             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 367             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 368
 369             /**************************
 370              * CALCULATE INTERACTIONS *
 371              **************************/
 372
 373             if (gmx_mm_any_lt(rsq11,rcutoff2))
 374             {
 375
 376             /* REACTION-FIELD ELECTROSTATICS */
 377             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 378             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 379
 380             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 381
 382             /* Update potential sum for this i atom from the interaction with this j atom. */
 383             velec            = _mm_and_pd(velec,cutoff_mask);
 384             velecsum         = _mm_add_pd(velecsum,velec);
 385
 386             fscal            = felec;
 387
 388             fscal            = _mm_and_pd(fscal,cutoff_mask);
 389
 390             /* Update vectorial force */
 391             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 392             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 393             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 394
 395             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 396             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 397             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 398
 399             }
 400
 401             /**************************
 402              * CALCULATE INTERACTIONS *
 403              **************************/
 404
 405             if (gmx_mm_any_lt(rsq12,rcutoff2))
 406             {
 407
 408             /* REACTION-FIELD ELECTROSTATICS */
 409             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 410             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 411
 412             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 413
 414             /* Update potential sum for this i atom from the interaction with this j atom. */
 415             velec            = _mm_and_pd(velec,cutoff_mask);
 416             velecsum         = _mm_add_pd(velecsum,velec);
 417
 418             fscal            = felec;
 419
 420             fscal            = _mm_and_pd(fscal,cutoff_mask);
 421
 422             /* Update vectorial force */
 423             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 424             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 425             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 426
 427             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 428             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 429             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 430
 431             }
 432
 433             /**************************
 434              * CALCULATE INTERACTIONS *
 435              **************************/
 436
 437             if (gmx_mm_any_lt(rsq13,rcutoff2))
 438             {
 439
 440             /* REACTION-FIELD ELECTROSTATICS */
 441             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 442             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 443
 444             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 445
 446             /* Update potential sum for this i atom from the interaction with this j atom. */
 447             velec            = _mm_and_pd(velec,cutoff_mask);
 448             velecsum         = _mm_add_pd(velecsum,velec);
 449
 450             fscal            = felec;
 451
 452             fscal            = _mm_and_pd(fscal,cutoff_mask);
 453
 454             /* Update vectorial force */
 455             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 456             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 457             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 458
 459             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 460             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 461             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 462
 463             }
 464
 465             /**************************
 466              * CALCULATE INTERACTIONS *
 467              **************************/
 468
 469             if (gmx_mm_any_lt(rsq21,rcutoff2))
 470             {
 471
 472             /* REACTION-FIELD ELECTROSTATICS */
 473             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 474             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 475
 476             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 477
 478             /* Update potential sum for this i atom from the interaction with this j atom. */
 479             velec            = _mm_and_pd(velec,cutoff_mask);
 480             velecsum         = _mm_add_pd(velecsum,velec);
 481
 482             fscal            = felec;
 483
 484             fscal            = _mm_and_pd(fscal,cutoff_mask);
 485
 486             /* Update vectorial force */
 487             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 488             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 489             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 490
 491             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 492             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 493             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 494
 495             }
 496
 497             /**************************
 498              * CALCULATE INTERACTIONS *
 499              **************************/
 500
 501             if (gmx_mm_any_lt(rsq22,rcutoff2))
 502             {
 503
 504             /* REACTION-FIELD ELECTROSTATICS */
 505             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 506             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 507
 508             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 509
 510             /* Update potential sum for this i atom from the interaction with this j atom. */
 511             velec            = _mm_and_pd(velec,cutoff_mask);
 512             velecsum         = _mm_add_pd(velecsum,velec);
 513
 514             fscal            = felec;
 515
 516             fscal            = _mm_and_pd(fscal,cutoff_mask);
 517
 518             /* Update vectorial force */
 519             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 520             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 521             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 522
 523             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 524             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 525             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 526
 527             }
 528
 529             /**************************
 530              * CALCULATE INTERACTIONS *
 531              **************************/
 532
 533             if (gmx_mm_any_lt(rsq23,rcutoff2))
 534             {
 535
 536             /* REACTION-FIELD ELECTROSTATICS */
 537             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 538             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 539
 540             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 541
 542             /* Update potential sum for this i atom from the interaction with this j atom. */
 543             velec            = _mm_and_pd(velec,cutoff_mask);
 544             velecsum         = _mm_add_pd(velecsum,velec);
 545
 546             fscal            = felec;
 547
 548             fscal            = _mm_and_pd(fscal,cutoff_mask);
 549
 550             /* Update vectorial force */
 551             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 552             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 553             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 554
 555             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 556             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 557             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 558
 559             }
 560
 561             /**************************
 562              * CALCULATE INTERACTIONS *
 563              **************************/
 564
 565             if (gmx_mm_any_lt(rsq31,rcutoff2))
 566             {
 567
 568             /* REACTION-FIELD ELECTROSTATICS */
 569             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
 570             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
 571
 572             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
 573
 574             /* Update potential sum for this i atom from the interaction with this j atom. */
 575             velec            = _mm_and_pd(velec,cutoff_mask);
 576             velecsum         = _mm_add_pd(velecsum,velec);
 577
 578             fscal            = felec;
 579
 580             fscal            = _mm_and_pd(fscal,cutoff_mask);
 581
 582             /* Update vectorial force */
 583             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 584             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 585             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 586
 587             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 588             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 589             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 590
 591             }
 592
 593             /**************************
 594              * CALCULATE INTERACTIONS *
 595              **************************/
 596
 597             if (gmx_mm_any_lt(rsq32,rcutoff2))
 598             {
 599
 600             /* REACTION-FIELD ELECTROSTATICS */
 601             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
 602             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
 603
 604             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
 605
 606             /* Update potential sum for this i atom from the interaction with this j atom. */
 607             velec            = _mm_and_pd(velec,cutoff_mask);
 608             velecsum         = _mm_add_pd(velecsum,velec);
 609
 610             fscal            = felec;
 611
 612             fscal            = _mm_and_pd(fscal,cutoff_mask);
 613
 614             /* Update vectorial force */
 615             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 616             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 617             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 618
 619             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 620             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 621             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 622
 623             }
 624
 625             /**************************
 626              * CALCULATE INTERACTIONS *
 627              **************************/
 628
 629             if (gmx_mm_any_lt(rsq33,rcutoff2))
 630             {
 631
 632             /* REACTION-FIELD ELECTROSTATICS */
 633             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
 634             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
 635
 636             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
 637
 638             /* Update potential sum for this i atom from the interaction with this j atom. */
 639             velec            = _mm_and_pd(velec,cutoff_mask);
 640             velecsum         = _mm_add_pd(velecsum,velec);
 641
 642             fscal            = felec;
 643
 644             fscal            = _mm_and_pd(fscal,cutoff_mask);
 645
 646             /* Update vectorial force */
 647             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 648             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 649             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 650
 651             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 652             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 653             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 654
 655             }
 656
 657             gmx_mm_decrement_4rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 658
 659             /* Inner loop uses 413 flops */
 660         }
 661
 662         if(jidx<j_index_end)
 663         {
 664
 665             jnrA             = jjnr[jidx];
 666             j_coord_offsetA  = DIM*jnrA;
 667
 668             /* load j atom coordinates */
 669             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 670                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 671                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 672
 673             /* Calculate displacement vector */
 674             dx00             = _mm_sub_pd(ix0,jx0);
 675             dy00             = _mm_sub_pd(iy0,jy0);
 676             dz00             = _mm_sub_pd(iz0,jz0);
 677             dx11             = _mm_sub_pd(ix1,jx1);
 678             dy11             = _mm_sub_pd(iy1,jy1);
 679             dz11             = _mm_sub_pd(iz1,jz1);
 680             dx12             = _mm_sub_pd(ix1,jx2);
 681             dy12             = _mm_sub_pd(iy1,jy2);
 682             dz12             = _mm_sub_pd(iz1,jz2);
 683             dx13             = _mm_sub_pd(ix1,jx3);
 684             dy13             = _mm_sub_pd(iy1,jy3);
 685             dz13             = _mm_sub_pd(iz1,jz3);
 686             dx21             = _mm_sub_pd(ix2,jx1);
 687             dy21             = _mm_sub_pd(iy2,jy1);
 688             dz21             = _mm_sub_pd(iz2,jz1);
 689             dx22             = _mm_sub_pd(ix2,jx2);
 690             dy22             = _mm_sub_pd(iy2,jy2);
 691             dz22             = _mm_sub_pd(iz2,jz2);
 692             dx23             = _mm_sub_pd(ix2,jx3);
 693             dy23             = _mm_sub_pd(iy2,jy3);
 694             dz23             = _mm_sub_pd(iz2,jz3);
 695             dx31             = _mm_sub_pd(ix3,jx1);
 696             dy31             = _mm_sub_pd(iy3,jy1);
 697             dz31             = _mm_sub_pd(iz3,jz1);
 698             dx32             = _mm_sub_pd(ix3,jx2);
 699             dy32             = _mm_sub_pd(iy3,jy2);
 700             dz32             = _mm_sub_pd(iz3,jz2);
 701             dx33             = _mm_sub_pd(ix3,jx3);
 702             dy33             = _mm_sub_pd(iy3,jy3);
 703             dz33             = _mm_sub_pd(iz3,jz3);
 704
 705             /* Calculate squared distance and things based on it */
 706             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 707             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 708             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 709             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 710             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 711             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 712             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 713             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 714             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 715             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 716
 717             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 718             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 719             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 720             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 721             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 722             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 723             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 724             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 725             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 726             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 727
 728             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 729             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 730             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 731             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 732             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 733             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 734             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 735             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 736             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 737
 738             fjx0             = _mm_setzero_pd();
 739             fjy0             = _mm_setzero_pd();
 740             fjz0             = _mm_setzero_pd();
 741             fjx1             = _mm_setzero_pd();
 742             fjy1             = _mm_setzero_pd();
 743             fjz1             = _mm_setzero_pd();
 744             fjx2             = _mm_setzero_pd();
 745             fjy2             = _mm_setzero_pd();
 746             fjz2             = _mm_setzero_pd();
 747             fjx3             = _mm_setzero_pd();
 748             fjy3             = _mm_setzero_pd();
 749             fjz3             = _mm_setzero_pd();
 750
 751             /**************************
 752              * CALCULATE INTERACTIONS *
 753              **************************/
 754
 755             r00              = _mm_mul_pd(rsq00,rinv00);
 756
 757             /* Calculate table index by multiplying r with table scale and truncate to integer */
 758             rt               = _mm_mul_pd(r00,vftabscale);
 759             vfitab           = _mm_cvttpd_epi32(rt);
 760 #ifdef __XOP__
 761             vfeps            = _mm_frcz_pd(rt);
 762 #else
 763             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 764 #endif
 765             twovfeps         = _mm_add_pd(vfeps,vfeps);
 766             vfitab           = _mm_slli_epi32(vfitab,3);
 767
 768             /* CUBIC SPLINE TABLE DISPERSION */
 769             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 770             F                = _mm_setzero_pd();
 771             GMX_MM_TRANSPOSE2_PD(Y,F);
 772             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 773             H                = _mm_setzero_pd();
 774             GMX_MM_TRANSPOSE2_PD(G,H);
 775             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 776             VV               = _mm_macc_pd(vfeps,Fp,Y);
 777             vvdw6            = _mm_mul_pd(c6_00,VV);
 778             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 779             fvdw6            = _mm_mul_pd(c6_00,FF);
 780
 781             /* CUBIC SPLINE TABLE REPULSION */
 782             vfitab           = _mm_add_epi32(vfitab,ifour);
 783             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 784             F                = _mm_setzero_pd();
 785             GMX_MM_TRANSPOSE2_PD(Y,F);
 786             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 787             H                = _mm_setzero_pd();
 788             GMX_MM_TRANSPOSE2_PD(G,H);
 789             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 790             VV               = _mm_macc_pd(vfeps,Fp,Y);
 791             vvdw12           = _mm_mul_pd(c12_00,VV);
 792             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 793             fvdw12           = _mm_mul_pd(c12_00,FF);
 794             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 795             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 796
 797             /* Update potential sum for this i atom from the interaction with this j atom. */
 798             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 799             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 800
 801             fscal            = fvdw;
 802
 803             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 804
 805             /* Update vectorial force */
 806             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 807             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 808             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 809
 810             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 811             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 812             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 813
 814             /**************************
 815              * CALCULATE INTERACTIONS *
 816              **************************/
 817
 818             if (gmx_mm_any_lt(rsq11,rcutoff2))
 819             {
 820
 821             /* REACTION-FIELD ELECTROSTATICS */
 822             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 823             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 824
 825             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 826
 827             /* Update potential sum for this i atom from the interaction with this j atom. */
 828             velec            = _mm_and_pd(velec,cutoff_mask);
 829             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 830             velecsum         = _mm_add_pd(velecsum,velec);
 831
 832             fscal            = felec;
 833
 834             fscal            = _mm_and_pd(fscal,cutoff_mask);
 835
 836             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 837
 838             /* Update vectorial force */
 839             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 840             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 841             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 842
 843             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 844             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 845             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 846
 847             }
 848
 849             /**************************
 850              * CALCULATE INTERACTIONS *
 851              **************************/
 852
 853             if (gmx_mm_any_lt(rsq12,rcutoff2))
 854             {
 855
 856             /* REACTION-FIELD ELECTROSTATICS */
 857             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 858             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 859
 860             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 861
 862             /* Update potential sum for this i atom from the interaction with this j atom. */
 863             velec            = _mm_and_pd(velec,cutoff_mask);
 864             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 865             velecsum         = _mm_add_pd(velecsum,velec);
 866
 867             fscal            = felec;
 868
 869             fscal            = _mm_and_pd(fscal,cutoff_mask);
 870
 871             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 872
 873             /* Update vectorial force */
 874             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 875             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 876             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 877
 878             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 879             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 880             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 881
 882             }
 883
 884             /**************************
 885              * CALCULATE INTERACTIONS *
 886              **************************/
 887
 888             if (gmx_mm_any_lt(rsq13,rcutoff2))
 889             {
 890
 891             /* REACTION-FIELD ELECTROSTATICS */
 892             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 893             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 894
 895             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 896
 897             /* Update potential sum for this i atom from the interaction with this j atom. */
 898             velec            = _mm_and_pd(velec,cutoff_mask);
 899             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 900             velecsum         = _mm_add_pd(velecsum,velec);
 901
 902             fscal            = felec;
 903
 904             fscal            = _mm_and_pd(fscal,cutoff_mask);
 905
 906             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 907
 908             /* Update vectorial force */
 909             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 910             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 911             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 912
 913             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 914             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 915             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 916
 917             }
 918
 919             /**************************
 920              * CALCULATE INTERACTIONS *
 921              **************************/
 922
 923             if (gmx_mm_any_lt(rsq21,rcutoff2))
 924             {
 925
 926             /* REACTION-FIELD ELECTROSTATICS */
 927             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 928             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 929
 930             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 931
 932             /* Update potential sum for this i atom from the interaction with this j atom. */
 933             velec            = _mm_and_pd(velec,cutoff_mask);
 934             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 935             velecsum         = _mm_add_pd(velecsum,velec);
 936
 937             fscal            = felec;
 938
 939             fscal            = _mm_and_pd(fscal,cutoff_mask);
 940
 941             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 942
 943             /* Update vectorial force */
 944             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 945             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 946             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 947
 948             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 949             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 950             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 951
 952             }
 953
 954             /**************************
 955              * CALCULATE INTERACTIONS *
 956              **************************/
 957
 958             if (gmx_mm_any_lt(rsq22,rcutoff2))
 959             {
 960
 961             /* REACTION-FIELD ELECTROSTATICS */
 962             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 963             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 964
 965             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 966
 967             /* Update potential sum for this i atom from the interaction with this j atom. */
 968             velec            = _mm_and_pd(velec,cutoff_mask);
 969             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 970             velecsum         = _mm_add_pd(velecsum,velec);
 971
 972             fscal            = felec;
 973
 974             fscal            = _mm_and_pd(fscal,cutoff_mask);
 975
 976             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 977
 978             /* Update vectorial force */
 979             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 980             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 981             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 982
 983             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 984             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 985             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 986
 987             }
 988
 989             /**************************
 990              * CALCULATE INTERACTIONS *
 991              **************************/
 992
 993             if (gmx_mm_any_lt(rsq23,rcutoff2))
 994             {
 995
 996             /* REACTION-FIELD ELECTROSTATICS */
 997             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 998             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 999
1000             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1001
1002             /* Update potential sum for this i atom from the interaction with this j atom. */
1003             velec            = _mm_and_pd(velec,cutoff_mask);
1004             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1005             velecsum         = _mm_add_pd(velecsum,velec);
1006
1007             fscal            = felec;
1008
1009             fscal            = _mm_and_pd(fscal,cutoff_mask);
1010
1011             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1012
1013             /* Update vectorial force */
1014             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1015             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1016             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1017
1018             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1019             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1020             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1021
1022             }
1023
1024             /**************************
1025              * CALCULATE INTERACTIONS *
1026              **************************/
1027
1028             if (gmx_mm_any_lt(rsq31,rcutoff2))
1029             {
1030
1031             /* REACTION-FIELD ELECTROSTATICS */
1032             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
1033             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1034
1035             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1036
1037             /* Update potential sum for this i atom from the interaction with this j atom. */
1038             velec            = _mm_and_pd(velec,cutoff_mask);
1039             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1040             velecsum         = _mm_add_pd(velecsum,velec);
1041
1042             fscal            = felec;
1043
1044             fscal            = _mm_and_pd(fscal,cutoff_mask);
1045
1046             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1047
1048             /* Update vectorial force */
1049             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1050             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1051             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1052
1053             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1054             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1055             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1056
1057             }
1058
1059             /**************************
1060              * CALCULATE INTERACTIONS *
1061              **************************/
1062
1063             if (gmx_mm_any_lt(rsq32,rcutoff2))
1064             {
1065
1066             /* REACTION-FIELD ELECTROSTATICS */
1067             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
1068             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1069
1070             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1071
1072             /* Update potential sum for this i atom from the interaction with this j atom. */
1073             velec            = _mm_and_pd(velec,cutoff_mask);
1074             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1075             velecsum         = _mm_add_pd(velecsum,velec);
1076
1077             fscal            = felec;
1078
1079             fscal            = _mm_and_pd(fscal,cutoff_mask);
1080
1081             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1082
1083             /* Update vectorial force */
1084             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1085             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1086             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1087
1088             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1089             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1090             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1091
1092             }
1093
1094             /**************************
1095              * CALCULATE INTERACTIONS *
1096              **************************/
1097
1098             if (gmx_mm_any_lt(rsq33,rcutoff2))
1099             {
1100
1101             /* REACTION-FIELD ELECTROSTATICS */
1102             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
1103             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1104
1105             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1106
1107             /* Update potential sum for this i atom from the interaction with this j atom. */
1108             velec            = _mm_and_pd(velec,cutoff_mask);
1109             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1110             velecsum         = _mm_add_pd(velecsum,velec);
1111
1112             fscal            = felec;
1113
1114             fscal            = _mm_and_pd(fscal,cutoff_mask);
1115
1116             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1117
1118             /* Update vectorial force */
1119             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1120             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1121             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1122
1123             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1124             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1125             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1126
1127             }
1128
1129             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1130
1131             /* Inner loop uses 413 flops */
1132         }
1133
1134         /* End of innermost loop */
1135
1136         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1137                                               f+i_coord_offset,fshift+i_shift_offset);
1138
1139         ggid                        = gid[iidx];
1140         /* Update potential energies */
1141         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1142         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
1143
1144         /* Increment number of inner iterations */
1145         inneriter                  += j_index_end - j_index_start;
1146
1147         /* Outer loop uses 26 flops */
1148     }
1149
1150     /* Increment number of outer iterations */
1151     outeriter        += nri;
1152
1153     /* Update outer/inner flops */
1154
1155     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*413);
1156 }
1157 /*
1158  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW4W4_F_avx_128_fma_double
1159  * Electrostatics interaction: ReactionField
1160  * VdW interaction:            CubicSplineTable
1161  * Geometry:                   Water4-Water4
1162  * Calculate force/pot:        Force
1163  */
1164 void
1165 nb_kernel_ElecRFCut_VdwCSTab_GeomW4W4_F_avx_128_fma_double
1166                     (t_nblist                    * gmx_restrict       nlist,
1167                      rvec                        * gmx_restrict          xx,
1168                      rvec                        * gmx_restrict          ff,
1169                      t_forcerec                  * gmx_restrict          fr,
1170                      t_mdatoms                   * gmx_restrict     mdatoms,
1171                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1172                      t_nrnb                      * gmx_restrict        nrnb)
1173 {
1174     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1175      * just 0 for non-waters.
1176      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1177      * jnr indices corresponding to data put in the four positions in the SIMD register.
1178      */
1179     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1180     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1181     int              jnrA,jnrB;
1182     int              j_coord_offsetA,j_coord_offsetB;
1183     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1184     real             rcutoff_scalar;
1185     real             *shiftvec,*fshift,*x,*f;
1186     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1187     int              vdwioffset0;
1188     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1189     int              vdwioffset1;
1190     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1191     int              vdwioffset2;
1192     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1193     int              vdwioffset3;
1194     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1195     int              vdwjidx0A,vdwjidx0B;
1196     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1197     int              vdwjidx1A,vdwjidx1B;
1198     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1199     int              vdwjidx2A,vdwjidx2B;
1200     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1201     int              vdwjidx3A,vdwjidx3B;
1202     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1203     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1204     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1205     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1206     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1207     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1208     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1209     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1210     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1211     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1212     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1213     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1214     real             *charge;
1215     int              nvdwtype;
1216     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1217     int              *vdwtype;
1218     real             *vdwparam;
1219     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1220     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1221     __m128i          vfitab;
1222     __m128i          ifour       = _mm_set1_epi32(4);
1223     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
1224     real             *vftab;
1225     __m128d          dummy_mask,cutoff_mask;
1226     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1227     __m128d          one     = _mm_set1_pd(1.0);
1228     __m128d          two     = _mm_set1_pd(2.0);
1229     x                = xx[0];
1230     f                = ff[0];
1231
1232     nri              = nlist->nri;
1233     iinr             = nlist->iinr;
1234     jindex           = nlist->jindex;
1235     jjnr             = nlist->jjnr;
1236     shiftidx         = nlist->shift;
1237     gid              = nlist->gid;
1238     shiftvec         = fr->shift_vec[0];
1239     fshift           = fr->fshift[0];
1240     facel            = _mm_set1_pd(fr->epsfac);
1241     charge           = mdatoms->chargeA;
1242     krf              = _mm_set1_pd(fr->ic->k_rf);
1243     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1244     crf              = _mm_set1_pd(fr->ic->c_rf);
1245     nvdwtype         = fr->ntype;
1246     vdwparam         = fr->nbfp;
1247     vdwtype          = mdatoms->typeA;
1248
1249     vftab            = kernel_data->table_vdw->data;
1250     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
1251
1252     /* Setup water-specific parameters */
1253     inr              = nlist->iinr[0];
1254     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1255     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1256     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1257     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1258
1259     jq1              = _mm_set1_pd(charge[inr+1]);
1260     jq2              = _mm_set1_pd(charge[inr+2]);
1261     jq3              = _mm_set1_pd(charge[inr+3]);
1262     vdwjidx0A        = 2*vdwtype[inr+0];
1263     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1264     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1265     qq11             = _mm_mul_pd(iq1,jq1);
1266     qq12             = _mm_mul_pd(iq1,jq2);
1267     qq13             = _mm_mul_pd(iq1,jq3);
1268     qq21             = _mm_mul_pd(iq2,jq1);
1269     qq22             = _mm_mul_pd(iq2,jq2);
1270     qq23             = _mm_mul_pd(iq2,jq3);
1271     qq31             = _mm_mul_pd(iq3,jq1);
1272     qq32             = _mm_mul_pd(iq3,jq2);
1273     qq33             = _mm_mul_pd(iq3,jq3);
1274
1275     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1276     rcutoff_scalar   = fr->rcoulomb;
1277     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1278     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1279
1280     /* Avoid stupid compiler warnings */
1281     jnrA = jnrB = 0;
1282     j_coord_offsetA = 0;
1283     j_coord_offsetB = 0;
1284
1285     outeriter        = 0;
1286     inneriter        = 0;
1287
1288     /* Start outer loop over neighborlists */
1289     for(iidx=0; iidx<nri; iidx++)
1290     {
1291         /* Load shift vector for this list */
1292         i_shift_offset   = DIM*shiftidx[iidx];
1293
1294         /* Load limits for loop over neighbors */
1295         j_index_start    = jindex[iidx];
1296         j_index_end      = jindex[iidx+1];
1297
1298         /* Get outer coordinate index */
1299         inr              = iinr[iidx];
1300         i_coord_offset   = DIM*inr;
1301
1302         /* Load i particle coords and add shift vector */
1303         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1304                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1305
1306         fix0             = _mm_setzero_pd();
1307         fiy0             = _mm_setzero_pd();
1308         fiz0             = _mm_setzero_pd();
1309         fix1             = _mm_setzero_pd();
1310         fiy1             = _mm_setzero_pd();
1311         fiz1             = _mm_setzero_pd();
1312         fix2             = _mm_setzero_pd();
1313         fiy2             = _mm_setzero_pd();
1314         fiz2             = _mm_setzero_pd();
1315         fix3             = _mm_setzero_pd();
1316         fiy3             = _mm_setzero_pd();
1317         fiz3             = _mm_setzero_pd();
1318
1319         /* Start inner kernel loop */
1320         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1321         {
1322
1323             /* Get j neighbor index, and coordinate index */
1324             jnrA             = jjnr[jidx];
1325             jnrB             = jjnr[jidx+1];
1326             j_coord_offsetA  = DIM*jnrA;
1327             j_coord_offsetB  = DIM*jnrB;
1328
1329             /* load j atom coordinates */
1330             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1331                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1332                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1333
1334             /* Calculate displacement vector */
1335             dx00             = _mm_sub_pd(ix0,jx0);
1336             dy00             = _mm_sub_pd(iy0,jy0);
1337             dz00             = _mm_sub_pd(iz0,jz0);
1338             dx11             = _mm_sub_pd(ix1,jx1);
1339             dy11             = _mm_sub_pd(iy1,jy1);
1340             dz11             = _mm_sub_pd(iz1,jz1);
1341             dx12             = _mm_sub_pd(ix1,jx2);
1342             dy12             = _mm_sub_pd(iy1,jy2);
1343             dz12             = _mm_sub_pd(iz1,jz2);
1344             dx13             = _mm_sub_pd(ix1,jx3);
1345             dy13             = _mm_sub_pd(iy1,jy3);
1346             dz13             = _mm_sub_pd(iz1,jz3);
1347             dx21             = _mm_sub_pd(ix2,jx1);
1348             dy21             = _mm_sub_pd(iy2,jy1);
1349             dz21             = _mm_sub_pd(iz2,jz1);
1350             dx22             = _mm_sub_pd(ix2,jx2);
1351             dy22             = _mm_sub_pd(iy2,jy2);
1352             dz22             = _mm_sub_pd(iz2,jz2);
1353             dx23             = _mm_sub_pd(ix2,jx3);
1354             dy23             = _mm_sub_pd(iy2,jy3);
1355             dz23             = _mm_sub_pd(iz2,jz3);
1356             dx31             = _mm_sub_pd(ix3,jx1);
1357             dy31             = _mm_sub_pd(iy3,jy1);
1358             dz31             = _mm_sub_pd(iz3,jz1);
1359             dx32             = _mm_sub_pd(ix3,jx2);
1360             dy32             = _mm_sub_pd(iy3,jy2);
1361             dz32             = _mm_sub_pd(iz3,jz2);
1362             dx33             = _mm_sub_pd(ix3,jx3);
1363             dy33             = _mm_sub_pd(iy3,jy3);
1364             dz33             = _mm_sub_pd(iz3,jz3);
1365
1366             /* Calculate squared distance and things based on it */
1367             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1368             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1369             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1370             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1371             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1372             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1373             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1374             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1375             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1376             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1377
1378             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1379             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1380             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1381             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1382             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1383             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1384             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1385             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1386             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1387             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1388
1389             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1390             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1391             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1392             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1393             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1394             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1395             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1396             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1397             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1398
1399             fjx0             = _mm_setzero_pd();
1400             fjy0             = _mm_setzero_pd();
1401             fjz0             = _mm_setzero_pd();
1402             fjx1             = _mm_setzero_pd();
1403             fjy1             = _mm_setzero_pd();
1404             fjz1             = _mm_setzero_pd();
1405             fjx2             = _mm_setzero_pd();
1406             fjy2             = _mm_setzero_pd();
1407             fjz2             = _mm_setzero_pd();
1408             fjx3             = _mm_setzero_pd();
1409             fjy3             = _mm_setzero_pd();
1410             fjz3             = _mm_setzero_pd();
1411
1412             /**************************
1413              * CALCULATE INTERACTIONS *
1414              **************************/
1415
1416             r00              = _mm_mul_pd(rsq00,rinv00);
1417
1418             /* Calculate table index by multiplying r with table scale and truncate to integer */
1419             rt               = _mm_mul_pd(r00,vftabscale);
1420             vfitab           = _mm_cvttpd_epi32(rt);
1421 #ifdef __XOP__
1422             vfeps            = _mm_frcz_pd(rt);
1423 #else
1424             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
1425 #endif
1426             twovfeps         = _mm_add_pd(vfeps,vfeps);
1427             vfitab           = _mm_slli_epi32(vfitab,3);
1428
1429             /* CUBIC SPLINE TABLE DISPERSION */
1430             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1431             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
1432             GMX_MM_TRANSPOSE2_PD(Y,F);
1433             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1434             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
1435             GMX_MM_TRANSPOSE2_PD(G,H);
1436             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1437             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1438             fvdw6            = _mm_mul_pd(c6_00,FF);
1439
1440             /* CUBIC SPLINE TABLE REPULSION */
1441             vfitab           = _mm_add_epi32(vfitab,ifour);
1442             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1443             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
1444             GMX_MM_TRANSPOSE2_PD(Y,F);
1445             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1446             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
1447             GMX_MM_TRANSPOSE2_PD(G,H);
1448             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1449             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1450             fvdw12           = _mm_mul_pd(c12_00,FF);
1451             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1452
1453             fscal            = fvdw;
1454
1455             /* Update vectorial force */
1456             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1457             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1458             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1459
1460             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1461             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1462             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1463
1464             /**************************
1465              * CALCULATE INTERACTIONS *
1466              **************************/
1467
1468             if (gmx_mm_any_lt(rsq11,rcutoff2))
1469             {
1470
1471             /* REACTION-FIELD ELECTROSTATICS */
1472             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1473
1474             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1475
1476             fscal            = felec;
1477
1478             fscal            = _mm_and_pd(fscal,cutoff_mask);
1479
1480             /* Update vectorial force */
1481             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1482             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1483             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1484
1485             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1486             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1487             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1488
1489             }
1490
1491             /**************************
1492              * CALCULATE INTERACTIONS *
1493              **************************/
1494
1495             if (gmx_mm_any_lt(rsq12,rcutoff2))
1496             {
1497
1498             /* REACTION-FIELD ELECTROSTATICS */
1499             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1500
1501             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1502
1503             fscal            = felec;
1504
1505             fscal            = _mm_and_pd(fscal,cutoff_mask);
1506
1507             /* Update vectorial force */
1508             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1509             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1510             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1511
1512             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1513             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1514             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1515
1516             }
1517
1518             /**************************
1519              * CALCULATE INTERACTIONS *
1520              **************************/
1521
1522             if (gmx_mm_any_lt(rsq13,rcutoff2))
1523             {
1524
1525             /* REACTION-FIELD ELECTROSTATICS */
1526             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1527
1528             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1529
1530             fscal            = felec;
1531
1532             fscal            = _mm_and_pd(fscal,cutoff_mask);
1533
1534             /* Update vectorial force */
1535             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1536             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1537             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1538
1539             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1540             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1541             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1542
1543             }
1544
1545             /**************************
1546              * CALCULATE INTERACTIONS *
1547              **************************/
1548
1549             if (gmx_mm_any_lt(rsq21,rcutoff2))
1550             {
1551
1552             /* REACTION-FIELD ELECTROSTATICS */
1553             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1554
1555             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1556
1557             fscal            = felec;
1558
1559             fscal            = _mm_and_pd(fscal,cutoff_mask);
1560
1561             /* Update vectorial force */
1562             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1563             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1564             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1565
1566             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1567             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1568             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1569
1570             }
1571
1572             /**************************
1573              * CALCULATE INTERACTIONS *
1574              **************************/
1575
1576             if (gmx_mm_any_lt(rsq22,rcutoff2))
1577             {
1578
1579             /* REACTION-FIELD ELECTROSTATICS */
1580             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1581
1582             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1583
1584             fscal            = felec;
1585
1586             fscal            = _mm_and_pd(fscal,cutoff_mask);
1587
1588             /* Update vectorial force */
1589             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1590             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1591             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1592
1593             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1594             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1595             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1596
1597             }
1598
1599             /**************************
1600              * CALCULATE INTERACTIONS *
1601              **************************/
1602
1603             if (gmx_mm_any_lt(rsq23,rcutoff2))
1604             {
1605
1606             /* REACTION-FIELD ELECTROSTATICS */
1607             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1608
1609             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1610
1611             fscal            = felec;
1612
1613             fscal            = _mm_and_pd(fscal,cutoff_mask);
1614
1615             /* Update vectorial force */
1616             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1617             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1618             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1619
1620             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1621             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1622             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1623
1624             }
1625
1626             /**************************
1627              * CALCULATE INTERACTIONS *
1628              **************************/
1629
1630             if (gmx_mm_any_lt(rsq31,rcutoff2))
1631             {
1632
1633             /* REACTION-FIELD ELECTROSTATICS */
1634             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1635
1636             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1637
1638             fscal            = felec;
1639
1640             fscal            = _mm_and_pd(fscal,cutoff_mask);
1641
1642             /* Update vectorial force */
1643             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1644             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1645             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1646
1647             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1648             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1649             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1650
1651             }
1652
1653             /**************************
1654              * CALCULATE INTERACTIONS *
1655              **************************/
1656
1657             if (gmx_mm_any_lt(rsq32,rcutoff2))
1658             {
1659
1660             /* REACTION-FIELD ELECTROSTATICS */
1661             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1662
1663             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1664
1665             fscal            = felec;
1666
1667             fscal            = _mm_and_pd(fscal,cutoff_mask);
1668
1669             /* Update vectorial force */
1670             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1671             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1672             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1673
1674             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1675             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1676             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1677
1678             }
1679
1680             /**************************
1681              * CALCULATE INTERACTIONS *
1682              **************************/
1683
1684             if (gmx_mm_any_lt(rsq33,rcutoff2))
1685             {
1686
1687             /* REACTION-FIELD ELECTROSTATICS */
1688             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1689
1690             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1691
1692             fscal            = felec;
1693
1694             fscal            = _mm_and_pd(fscal,cutoff_mask);
1695
1696             /* Update vectorial force */
1697             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1698             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1699             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1700
1701             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1702             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1703             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1704
1705             }
1706
1707             gmx_mm_decrement_4rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1708
1709             /* Inner loop uses 351 flops */
1710         }
1711
1712         if(jidx<j_index_end)
1713         {
1714
1715             jnrA             = jjnr[jidx];
1716             j_coord_offsetA  = DIM*jnrA;
1717
1718             /* load j atom coordinates */
1719             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1720                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1721                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1722
1723             /* Calculate displacement vector */
1724             dx00             = _mm_sub_pd(ix0,jx0);
1725             dy00             = _mm_sub_pd(iy0,jy0);
1726             dz00             = _mm_sub_pd(iz0,jz0);
1727             dx11             = _mm_sub_pd(ix1,jx1);
1728             dy11             = _mm_sub_pd(iy1,jy1);
1729             dz11             = _mm_sub_pd(iz1,jz1);
1730             dx12             = _mm_sub_pd(ix1,jx2);
1731             dy12             = _mm_sub_pd(iy1,jy2);
1732             dz12             = _mm_sub_pd(iz1,jz2);
1733             dx13             = _mm_sub_pd(ix1,jx3);
1734             dy13             = _mm_sub_pd(iy1,jy3);
1735             dz13             = _mm_sub_pd(iz1,jz3);
1736             dx21             = _mm_sub_pd(ix2,jx1);
1737             dy21             = _mm_sub_pd(iy2,jy1);
1738             dz21             = _mm_sub_pd(iz2,jz1);
1739             dx22             = _mm_sub_pd(ix2,jx2);
1740             dy22             = _mm_sub_pd(iy2,jy2);
1741             dz22             = _mm_sub_pd(iz2,jz2);
1742             dx23             = _mm_sub_pd(ix2,jx3);
1743             dy23             = _mm_sub_pd(iy2,jy3);
1744             dz23             = _mm_sub_pd(iz2,jz3);
1745             dx31             = _mm_sub_pd(ix3,jx1);
1746             dy31             = _mm_sub_pd(iy3,jy1);
1747             dz31             = _mm_sub_pd(iz3,jz1);
1748             dx32             = _mm_sub_pd(ix3,jx2);
1749             dy32             = _mm_sub_pd(iy3,jy2);
1750             dz32             = _mm_sub_pd(iz3,jz2);
1751             dx33             = _mm_sub_pd(ix3,jx3);
1752             dy33             = _mm_sub_pd(iy3,jy3);
1753             dz33             = _mm_sub_pd(iz3,jz3);
1754
1755             /* Calculate squared distance and things based on it */
1756             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1757             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1758             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1759             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1760             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1761             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1762             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1763             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1764             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1765             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1766
1767             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1768             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1769             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1770             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1771             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1772             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1773             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1774             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1775             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1776             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1777
1778             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1779             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1780             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1781             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1782             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1783             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1784             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1785             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1786             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1787
1788             fjx0             = _mm_setzero_pd();
1789             fjy0             = _mm_setzero_pd();
1790             fjz0             = _mm_setzero_pd();
1791             fjx1             = _mm_setzero_pd();
1792             fjy1             = _mm_setzero_pd();
1793             fjz1             = _mm_setzero_pd();
1794             fjx2             = _mm_setzero_pd();
1795             fjy2             = _mm_setzero_pd();
1796             fjz2             = _mm_setzero_pd();
1797             fjx3             = _mm_setzero_pd();
1798             fjy3             = _mm_setzero_pd();
1799             fjz3             = _mm_setzero_pd();
1800
1801             /**************************
1802              * CALCULATE INTERACTIONS *
1803              **************************/
1804
1805             r00              = _mm_mul_pd(rsq00,rinv00);
1806
1807             /* Calculate table index by multiplying r with table scale and truncate to integer */
1808             rt               = _mm_mul_pd(r00,vftabscale);
1809             vfitab           = _mm_cvttpd_epi32(rt);
1810 #ifdef __XOP__
1811             vfeps            = _mm_frcz_pd(rt);
1812 #else
1813             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
1814 #endif
1815             twovfeps         = _mm_add_pd(vfeps,vfeps);
1816             vfitab           = _mm_slli_epi32(vfitab,3);
1817
1818             /* CUBIC SPLINE TABLE DISPERSION */
1819             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1820             F                = _mm_setzero_pd();
1821             GMX_MM_TRANSPOSE2_PD(Y,F);
1822             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1823             H                = _mm_setzero_pd();
1824             GMX_MM_TRANSPOSE2_PD(G,H);
1825             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1826             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1827             fvdw6            = _mm_mul_pd(c6_00,FF);
1828
1829             /* CUBIC SPLINE TABLE REPULSION */
1830             vfitab           = _mm_add_epi32(vfitab,ifour);
1831             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1832             F                = _mm_setzero_pd();
1833             GMX_MM_TRANSPOSE2_PD(Y,F);
1834             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1835             H                = _mm_setzero_pd();
1836             GMX_MM_TRANSPOSE2_PD(G,H);
1837             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1838             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1839             fvdw12           = _mm_mul_pd(c12_00,FF);
1840             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1841
1842             fscal            = fvdw;
1843
1844             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1845
1846             /* Update vectorial force */
1847             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1848             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1849             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1850
1851             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1852             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1853             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1854
1855             /**************************
1856              * CALCULATE INTERACTIONS *
1857              **************************/
1858
1859             if (gmx_mm_any_lt(rsq11,rcutoff2))
1860             {
1861
1862             /* REACTION-FIELD ELECTROSTATICS */
1863             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1864
1865             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1866
1867             fscal            = felec;
1868
1869             fscal            = _mm_and_pd(fscal,cutoff_mask);
1870
1871             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1872
1873             /* Update vectorial force */
1874             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1875             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1876             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1877
1878             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1879             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1880             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1881
1882             }
1883
1884             /**************************
1885              * CALCULATE INTERACTIONS *
1886              **************************/
1887
1888             if (gmx_mm_any_lt(rsq12,rcutoff2))
1889             {
1890
1891             /* REACTION-FIELD ELECTROSTATICS */
1892             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1893
1894             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1895
1896             fscal            = felec;
1897
1898             fscal            = _mm_and_pd(fscal,cutoff_mask);
1899
1900             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1901
1902             /* Update vectorial force */
1903             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1904             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1905             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1906
1907             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1908             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1909             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1910
1911             }
1912
1913             /**************************
1914              * CALCULATE INTERACTIONS *
1915              **************************/
1916
1917             if (gmx_mm_any_lt(rsq13,rcutoff2))
1918             {
1919
1920             /* REACTION-FIELD ELECTROSTATICS */
1921             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1922
1923             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1924
1925             fscal            = felec;
1926
1927             fscal            = _mm_and_pd(fscal,cutoff_mask);
1928
1929             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1930
1931             /* Update vectorial force */
1932             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1933             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1934             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1935
1936             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1937             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1938             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1939
1940             }
1941
1942             /**************************
1943              * CALCULATE INTERACTIONS *
1944              **************************/
1945
1946             if (gmx_mm_any_lt(rsq21,rcutoff2))
1947             {
1948
1949             /* REACTION-FIELD ELECTROSTATICS */
1950             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1951
1952             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1953
1954             fscal            = felec;
1955
1956             fscal            = _mm_and_pd(fscal,cutoff_mask);
1957
1958             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1959
1960             /* Update vectorial force */
1961             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1962             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1963             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1964
1965             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1966             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1967             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1968
1969             }
1970
1971             /**************************
1972              * CALCULATE INTERACTIONS *
1973              **************************/
1974
1975             if (gmx_mm_any_lt(rsq22,rcutoff2))
1976             {
1977
1978             /* REACTION-FIELD ELECTROSTATICS */
1979             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1980
1981             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1982
1983             fscal            = felec;
1984
1985             fscal            = _mm_and_pd(fscal,cutoff_mask);
1986
1987             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1988
1989             /* Update vectorial force */
1990             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1991             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1992             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1993
1994             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1995             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1996             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1997
1998             }
1999
2000             /**************************
2001              * CALCULATE INTERACTIONS *
2002              **************************/
2003
2004             if (gmx_mm_any_lt(rsq23,rcutoff2))
2005             {
2006
2007             /* REACTION-FIELD ELECTROSTATICS */
2008             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
2009
2010             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
2011
2012             fscal            = felec;
2013
2014             fscal            = _mm_and_pd(fscal,cutoff_mask);
2015
2016             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2017
2018             /* Update vectorial force */
2019             fix2             = _mm_macc_pd(dx23,fscal,fix2);
2020             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
2021             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
2022
2023             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
2024             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
2025             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
2026
2027             }
2028
2029             /**************************
2030              * CALCULATE INTERACTIONS *
2031              **************************/
2032
2033             if (gmx_mm_any_lt(rsq31,rcutoff2))
2034             {
2035
2036             /* REACTION-FIELD ELECTROSTATICS */
2037             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
2038
2039             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
2040
2041             fscal            = felec;
2042
2043             fscal            = _mm_and_pd(fscal,cutoff_mask);
2044
2045             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2046
2047             /* Update vectorial force */
2048             fix3             = _mm_macc_pd(dx31,fscal,fix3);
2049             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
2050             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
2051
2052             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
2053             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
2054             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
2055
2056             }
2057
2058             /**************************
2059              * CALCULATE INTERACTIONS *
2060              **************************/
2061
2062             if (gmx_mm_any_lt(rsq32,rcutoff2))
2063             {
2064
2065             /* REACTION-FIELD ELECTROSTATICS */
2066             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
2067
2068             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
2069
2070             fscal            = felec;
2071
2072             fscal            = _mm_and_pd(fscal,cutoff_mask);
2073
2074             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2075
2076             /* Update vectorial force */
2077             fix3             = _mm_macc_pd(dx32,fscal,fix3);
2078             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
2079             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
2080
2081             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
2082             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
2083             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
2084
2085             }
2086
2087             /**************************
2088              * CALCULATE INTERACTIONS *
2089              **************************/
2090
2091             if (gmx_mm_any_lt(rsq33,rcutoff2))
2092             {
2093
2094             /* REACTION-FIELD ELECTROSTATICS */
2095             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
2096
2097             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
2098
2099             fscal            = felec;
2100
2101             fscal            = _mm_and_pd(fscal,cutoff_mask);
2102
2103             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2104
2105             /* Update vectorial force */
2106             fix3             = _mm_macc_pd(dx33,fscal,fix3);
2107             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
2108             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
2109
2110             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
2111             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
2112             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
2113
2114             }
2115
2116             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2117
2118             /* Inner loop uses 351 flops */
2119         }
2120
2121         /* End of innermost loop */
2122
2123         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2124                                               f+i_coord_offset,fshift+i_shift_offset);
2125
2126         /* Increment number of inner iterations */
2127         inneriter                  += j_index_end - j_index_start;
2128
2129         /* Outer loop uses 24 flops */
2130     }
2131
2132     /* Increment number of outer iterations */
2133     outeriter        += nri;
2134
2135     /* Update outer/inner flops */
2136
2137     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*351);
2138 }