src/gromacs/gmxlib/nonbonded/nb_kernel_sse4_1_double/nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_sse4_1_double.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse4_1_double kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_sse4_1_double.h"
  34 #include "kernelutil_x86_sse4_1_double.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_VF_sse4_1_double
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            CubicSplineTable
  40  * Geometry:                   Water3-Water3
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_VF_sse4_1_double
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB;
  61     int              j_coord_offsetA,j_coord_offsetB;
  62     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  63     real             rcutoff_scalar;
  64     real             *shiftvec,*fshift,*x,*f;
  65     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  66     int              vdwioffset0;
  67     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  68     int              vdwioffset1;
  69     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  70     int              vdwioffset2;
  71     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  72     int              vdwjidx0A,vdwjidx0B;
  73     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  74     int              vdwjidx1A,vdwjidx1B;
  75     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  76     int              vdwjidx2A,vdwjidx2B;
  77     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  78     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  79     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  80     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  81     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  82     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  83     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  84     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  85     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  86     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  87     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  88     real             *charge;
  89     int              nvdwtype;
  90     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  91     int              *vdwtype;
  92     real             *vdwparam;
  93     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
  94     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
  95     __m128i          vfitab;
  96     __m128i          ifour       = _mm_set1_epi32(4);
  97     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  98     real             *vftab;
  99     __m128d          dummy_mask,cutoff_mask;
 100     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 101     __m128d          one     = _mm_set1_pd(1.0);
 102     __m128d          two     = _mm_set1_pd(2.0);
 103     x                = xx[0];
 104     f                = ff[0];
 105
 106     nri              = nlist->nri;
 107     iinr             = nlist->iinr;
 108     jindex           = nlist->jindex;
 109     jjnr             = nlist->jjnr;
 110     shiftidx         = nlist->shift;
 111     gid              = nlist->gid;
 112     shiftvec         = fr->shift_vec[0];
 113     fshift           = fr->fshift[0];
 114     facel            = _mm_set1_pd(fr->epsfac);
 115     charge           = mdatoms->chargeA;
 116     krf              = _mm_set1_pd(fr->ic->k_rf);
 117     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 118     crf              = _mm_set1_pd(fr->ic->c_rf);
 119     nvdwtype         = fr->ntype;
 120     vdwparam         = fr->nbfp;
 121     vdwtype          = mdatoms->typeA;
 122
 123     vftab            = kernel_data->table_vdw->data;
 124     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
 125
 126     /* Setup water-specific parameters */
 127     inr              = nlist->iinr[0];
 128     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 129     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 130     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 131     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 132
 133     jq0              = _mm_set1_pd(charge[inr+0]);
 134     jq1              = _mm_set1_pd(charge[inr+1]);
 135     jq2              = _mm_set1_pd(charge[inr+2]);
 136     vdwjidx0A        = 2*vdwtype[inr+0];
 137     qq00             = _mm_mul_pd(iq0,jq0);
 138     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 139     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 140     qq01             = _mm_mul_pd(iq0,jq1);
 141     qq02             = _mm_mul_pd(iq0,jq2);
 142     qq10             = _mm_mul_pd(iq1,jq0);
 143     qq11             = _mm_mul_pd(iq1,jq1);
 144     qq12             = _mm_mul_pd(iq1,jq2);
 145     qq20             = _mm_mul_pd(iq2,jq0);
 146     qq21             = _mm_mul_pd(iq2,jq1);
 147     qq22             = _mm_mul_pd(iq2,jq2);
 148
 149     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 150     rcutoff_scalar   = fr->rcoulomb;
 151     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 152     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 153
 154     /* Avoid stupid compiler warnings */
 155     jnrA = jnrB = 0;
 156     j_coord_offsetA = 0;
 157     j_coord_offsetB = 0;
 158
 159     outeriter        = 0;
 160     inneriter        = 0;
 161
 162     /* Start outer loop over neighborlists */
 163     for(iidx=0; iidx<nri; iidx++)
 164     {
 165         /* Load shift vector for this list */
 166         i_shift_offset   = DIM*shiftidx[iidx];
 167
 168         /* Load limits for loop over neighbors */
 169         j_index_start    = jindex[iidx];
 170         j_index_end      = jindex[iidx+1];
 171
 172         /* Get outer coordinate index */
 173         inr              = iinr[iidx];
 174         i_coord_offset   = DIM*inr;
 175
 176         /* Load i particle coords and add shift vector */
 177         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 178                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 179
 180         fix0             = _mm_setzero_pd();
 181         fiy0             = _mm_setzero_pd();
 182         fiz0             = _mm_setzero_pd();
 183         fix1             = _mm_setzero_pd();
 184         fiy1             = _mm_setzero_pd();
 185         fiz1             = _mm_setzero_pd();
 186         fix2             = _mm_setzero_pd();
 187         fiy2             = _mm_setzero_pd();
 188         fiz2             = _mm_setzero_pd();
 189
 190         /* Reset potential sums */
 191         velecsum         = _mm_setzero_pd();
 192         vvdwsum          = _mm_setzero_pd();
 193
 194         /* Start inner kernel loop */
 195         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 196         {
 197
 198             /* Get j neighbor index, and coordinate index */
 199             jnrA             = jjnr[jidx];
 200             jnrB             = jjnr[jidx+1];
 201             j_coord_offsetA  = DIM*jnrA;
 202             j_coord_offsetB  = DIM*jnrB;
 203
 204             /* load j atom coordinates */
 205             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 206                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 207
 208             /* Calculate displacement vector */
 209             dx00             = _mm_sub_pd(ix0,jx0);
 210             dy00             = _mm_sub_pd(iy0,jy0);
 211             dz00             = _mm_sub_pd(iz0,jz0);
 212             dx01             = _mm_sub_pd(ix0,jx1);
 213             dy01             = _mm_sub_pd(iy0,jy1);
 214             dz01             = _mm_sub_pd(iz0,jz1);
 215             dx02             = _mm_sub_pd(ix0,jx2);
 216             dy02             = _mm_sub_pd(iy0,jy2);
 217             dz02             = _mm_sub_pd(iz0,jz2);
 218             dx10             = _mm_sub_pd(ix1,jx0);
 219             dy10             = _mm_sub_pd(iy1,jy0);
 220             dz10             = _mm_sub_pd(iz1,jz0);
 221             dx11             = _mm_sub_pd(ix1,jx1);
 222             dy11             = _mm_sub_pd(iy1,jy1);
 223             dz11             = _mm_sub_pd(iz1,jz1);
 224             dx12             = _mm_sub_pd(ix1,jx2);
 225             dy12             = _mm_sub_pd(iy1,jy2);
 226             dz12             = _mm_sub_pd(iz1,jz2);
 227             dx20             = _mm_sub_pd(ix2,jx0);
 228             dy20             = _mm_sub_pd(iy2,jy0);
 229             dz20             = _mm_sub_pd(iz2,jz0);
 230             dx21             = _mm_sub_pd(ix2,jx1);
 231             dy21             = _mm_sub_pd(iy2,jy1);
 232             dz21             = _mm_sub_pd(iz2,jz1);
 233             dx22             = _mm_sub_pd(ix2,jx2);
 234             dy22             = _mm_sub_pd(iy2,jy2);
 235             dz22             = _mm_sub_pd(iz2,jz2);
 236
 237             /* Calculate squared distance and things based on it */
 238             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 239             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 240             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 241             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 242             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 243             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 244             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 245             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 246             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 247
 248             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 249             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 250             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 251             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 252             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 253             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 254             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 255             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 256             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 257
 258             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 259             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 260             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 261             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 262             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 263             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 264             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 265             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 266             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 267
 268             fjx0             = _mm_setzero_pd();
 269             fjy0             = _mm_setzero_pd();
 270             fjz0             = _mm_setzero_pd();
 271             fjx1             = _mm_setzero_pd();
 272             fjy1             = _mm_setzero_pd();
 273             fjz1             = _mm_setzero_pd();
 274             fjx2             = _mm_setzero_pd();
 275             fjy2             = _mm_setzero_pd();
 276             fjz2             = _mm_setzero_pd();
 277
 278             /**************************
 279              * CALCULATE INTERACTIONS *
 280              **************************/
 281
 282             if (gmx_mm_any_lt(rsq00,rcutoff2))
 283             {
 284
 285             r00              = _mm_mul_pd(rsq00,rinv00);
 286
 287             /* Calculate table index by multiplying r with table scale and truncate to integer */
 288             rt               = _mm_mul_pd(r00,vftabscale);
 289             vfitab           = _mm_cvttpd_epi32(rt);
 290             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 291             vfitab           = _mm_slli_epi32(vfitab,3);
 292
 293             /* REACTION-FIELD ELECTROSTATICS */
 294             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
 295             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 296
 297             /* CUBIC SPLINE TABLE DISPERSION */
 298             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 299             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
 300             GMX_MM_TRANSPOSE2_PD(Y,F);
 301             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 302             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
 303             GMX_MM_TRANSPOSE2_PD(G,H);
 304             Heps             = _mm_mul_pd(vfeps,H);
 305             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 306             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 307             vvdw6            = _mm_mul_pd(c6_00,VV);
 308             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 309             fvdw6            = _mm_mul_pd(c6_00,FF);
 310
 311             /* CUBIC SPLINE TABLE REPULSION */
 312             vfitab           = _mm_add_epi32(vfitab,ifour);
 313             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 314             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
 315             GMX_MM_TRANSPOSE2_PD(Y,F);
 316             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 317             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
 318             GMX_MM_TRANSPOSE2_PD(G,H);
 319             Heps             = _mm_mul_pd(vfeps,H);
 320             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 321             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 322             vvdw12           = _mm_mul_pd(c12_00,VV);
 323             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 324             fvdw12           = _mm_mul_pd(c12_00,FF);
 325             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 326             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 327
 328             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 329
 330             /* Update potential sum for this i atom from the interaction with this j atom. */
 331             velec            = _mm_and_pd(velec,cutoff_mask);
 332             velecsum         = _mm_add_pd(velecsum,velec);
 333             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 334             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 335
 336             fscal            = _mm_add_pd(felec,fvdw);
 337
 338             fscal            = _mm_and_pd(fscal,cutoff_mask);
 339
 340             /* Calculate temporary vectorial force */
 341             tx               = _mm_mul_pd(fscal,dx00);
 342             ty               = _mm_mul_pd(fscal,dy00);
 343             tz               = _mm_mul_pd(fscal,dz00);
 344
 345             /* Update vectorial force */
 346             fix0             = _mm_add_pd(fix0,tx);
 347             fiy0             = _mm_add_pd(fiy0,ty);
 348             fiz0             = _mm_add_pd(fiz0,tz);
 349
 350             fjx0             = _mm_add_pd(fjx0,tx);
 351             fjy0             = _mm_add_pd(fjy0,ty);
 352             fjz0             = _mm_add_pd(fjz0,tz);
 353
 354             }
 355
 356             /**************************
 357              * CALCULATE INTERACTIONS *
 358              **************************/
 359
 360             if (gmx_mm_any_lt(rsq01,rcutoff2))
 361             {
 362
 363             /* REACTION-FIELD ELECTROSTATICS */
 364             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_add_pd(rinv01,_mm_mul_pd(krf,rsq01)),crf));
 365             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
 366
 367             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
 368
 369             /* Update potential sum for this i atom from the interaction with this j atom. */
 370             velec            = _mm_and_pd(velec,cutoff_mask);
 371             velecsum         = _mm_add_pd(velecsum,velec);
 372
 373             fscal            = felec;
 374
 375             fscal            = _mm_and_pd(fscal,cutoff_mask);
 376
 377             /* Calculate temporary vectorial force */
 378             tx               = _mm_mul_pd(fscal,dx01);
 379             ty               = _mm_mul_pd(fscal,dy01);
 380             tz               = _mm_mul_pd(fscal,dz01);
 381
 382             /* Update vectorial force */
 383             fix0             = _mm_add_pd(fix0,tx);
 384             fiy0             = _mm_add_pd(fiy0,ty);
 385             fiz0             = _mm_add_pd(fiz0,tz);
 386
 387             fjx1             = _mm_add_pd(fjx1,tx);
 388             fjy1             = _mm_add_pd(fjy1,ty);
 389             fjz1             = _mm_add_pd(fjz1,tz);
 390
 391             }
 392
 393             /**************************
 394              * CALCULATE INTERACTIONS *
 395              **************************/
 396
 397             if (gmx_mm_any_lt(rsq02,rcutoff2))
 398             {
 399
 400             /* REACTION-FIELD ELECTROSTATICS */
 401             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_add_pd(rinv02,_mm_mul_pd(krf,rsq02)),crf));
 402             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
 403
 404             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
 405
 406             /* Update potential sum for this i atom from the interaction with this j atom. */
 407             velec            = _mm_and_pd(velec,cutoff_mask);
 408             velecsum         = _mm_add_pd(velecsum,velec);
 409
 410             fscal            = felec;
 411
 412             fscal            = _mm_and_pd(fscal,cutoff_mask);
 413
 414             /* Calculate temporary vectorial force */
 415             tx               = _mm_mul_pd(fscal,dx02);
 416             ty               = _mm_mul_pd(fscal,dy02);
 417             tz               = _mm_mul_pd(fscal,dz02);
 418
 419             /* Update vectorial force */
 420             fix0             = _mm_add_pd(fix0,tx);
 421             fiy0             = _mm_add_pd(fiy0,ty);
 422             fiz0             = _mm_add_pd(fiz0,tz);
 423
 424             fjx2             = _mm_add_pd(fjx2,tx);
 425             fjy2             = _mm_add_pd(fjy2,ty);
 426             fjz2             = _mm_add_pd(fjz2,tz);
 427
 428             }
 429
 430             /**************************
 431              * CALCULATE INTERACTIONS *
 432              **************************/
 433
 434             if (gmx_mm_any_lt(rsq10,rcutoff2))
 435             {
 436
 437             /* REACTION-FIELD ELECTROSTATICS */
 438             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 439             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 440
 441             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
 442
 443             /* Update potential sum for this i atom from the interaction with this j atom. */
 444             velec            = _mm_and_pd(velec,cutoff_mask);
 445             velecsum         = _mm_add_pd(velecsum,velec);
 446
 447             fscal            = felec;
 448
 449             fscal            = _mm_and_pd(fscal,cutoff_mask);
 450
 451             /* Calculate temporary vectorial force */
 452             tx               = _mm_mul_pd(fscal,dx10);
 453             ty               = _mm_mul_pd(fscal,dy10);
 454             tz               = _mm_mul_pd(fscal,dz10);
 455
 456             /* Update vectorial force */
 457             fix1             = _mm_add_pd(fix1,tx);
 458             fiy1             = _mm_add_pd(fiy1,ty);
 459             fiz1             = _mm_add_pd(fiz1,tz);
 460
 461             fjx0             = _mm_add_pd(fjx0,tx);
 462             fjy0             = _mm_add_pd(fjy0,ty);
 463             fjz0             = _mm_add_pd(fjz0,tz);
 464
 465             }
 466
 467             /**************************
 468              * CALCULATE INTERACTIONS *
 469              **************************/
 470
 471             if (gmx_mm_any_lt(rsq11,rcutoff2))
 472             {
 473
 474             /* REACTION-FIELD ELECTROSTATICS */
 475             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 476             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 477
 478             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 479
 480             /* Update potential sum for this i atom from the interaction with this j atom. */
 481             velec            = _mm_and_pd(velec,cutoff_mask);
 482             velecsum         = _mm_add_pd(velecsum,velec);
 483
 484             fscal            = felec;
 485
 486             fscal            = _mm_and_pd(fscal,cutoff_mask);
 487
 488             /* Calculate temporary vectorial force */
 489             tx               = _mm_mul_pd(fscal,dx11);
 490             ty               = _mm_mul_pd(fscal,dy11);
 491             tz               = _mm_mul_pd(fscal,dz11);
 492
 493             /* Update vectorial force */
 494             fix1             = _mm_add_pd(fix1,tx);
 495             fiy1             = _mm_add_pd(fiy1,ty);
 496             fiz1             = _mm_add_pd(fiz1,tz);
 497
 498             fjx1             = _mm_add_pd(fjx1,tx);
 499             fjy1             = _mm_add_pd(fjy1,ty);
 500             fjz1             = _mm_add_pd(fjz1,tz);
 501
 502             }
 503
 504             /**************************
 505              * CALCULATE INTERACTIONS *
 506              **************************/
 507
 508             if (gmx_mm_any_lt(rsq12,rcutoff2))
 509             {
 510
 511             /* REACTION-FIELD ELECTROSTATICS */
 512             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 513             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 514
 515             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 516
 517             /* Update potential sum for this i atom from the interaction with this j atom. */
 518             velec            = _mm_and_pd(velec,cutoff_mask);
 519             velecsum         = _mm_add_pd(velecsum,velec);
 520
 521             fscal            = felec;
 522
 523             fscal            = _mm_and_pd(fscal,cutoff_mask);
 524
 525             /* Calculate temporary vectorial force */
 526             tx               = _mm_mul_pd(fscal,dx12);
 527             ty               = _mm_mul_pd(fscal,dy12);
 528             tz               = _mm_mul_pd(fscal,dz12);
 529
 530             /* Update vectorial force */
 531             fix1             = _mm_add_pd(fix1,tx);
 532             fiy1             = _mm_add_pd(fiy1,ty);
 533             fiz1             = _mm_add_pd(fiz1,tz);
 534
 535             fjx2             = _mm_add_pd(fjx2,tx);
 536             fjy2             = _mm_add_pd(fjy2,ty);
 537             fjz2             = _mm_add_pd(fjz2,tz);
 538
 539             }
 540
 541             /**************************
 542              * CALCULATE INTERACTIONS *
 543              **************************/
 544
 545             if (gmx_mm_any_lt(rsq20,rcutoff2))
 546             {
 547
 548             /* REACTION-FIELD ELECTROSTATICS */
 549             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
 550             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
 551
 552             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
 553
 554             /* Update potential sum for this i atom from the interaction with this j atom. */
 555             velec            = _mm_and_pd(velec,cutoff_mask);
 556             velecsum         = _mm_add_pd(velecsum,velec);
 557
 558             fscal            = felec;
 559
 560             fscal            = _mm_and_pd(fscal,cutoff_mask);
 561
 562             /* Calculate temporary vectorial force */
 563             tx               = _mm_mul_pd(fscal,dx20);
 564             ty               = _mm_mul_pd(fscal,dy20);
 565             tz               = _mm_mul_pd(fscal,dz20);
 566
 567             /* Update vectorial force */
 568             fix2             = _mm_add_pd(fix2,tx);
 569             fiy2             = _mm_add_pd(fiy2,ty);
 570             fiz2             = _mm_add_pd(fiz2,tz);
 571
 572             fjx0             = _mm_add_pd(fjx0,tx);
 573             fjy0             = _mm_add_pd(fjy0,ty);
 574             fjz0             = _mm_add_pd(fjz0,tz);
 575
 576             }
 577
 578             /**************************
 579              * CALCULATE INTERACTIONS *
 580              **************************/
 581
 582             if (gmx_mm_any_lt(rsq21,rcutoff2))
 583             {
 584
 585             /* REACTION-FIELD ELECTROSTATICS */
 586             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
 587             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
 588
 589             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 590
 591             /* Update potential sum for this i atom from the interaction with this j atom. */
 592             velec            = _mm_and_pd(velec,cutoff_mask);
 593             velecsum         = _mm_add_pd(velecsum,velec);
 594
 595             fscal            = felec;
 596
 597             fscal            = _mm_and_pd(fscal,cutoff_mask);
 598
 599             /* Calculate temporary vectorial force */
 600             tx               = _mm_mul_pd(fscal,dx21);
 601             ty               = _mm_mul_pd(fscal,dy21);
 602             tz               = _mm_mul_pd(fscal,dz21);
 603
 604             /* Update vectorial force */
 605             fix2             = _mm_add_pd(fix2,tx);
 606             fiy2             = _mm_add_pd(fiy2,ty);
 607             fiz2             = _mm_add_pd(fiz2,tz);
 608
 609             fjx1             = _mm_add_pd(fjx1,tx);
 610             fjy1             = _mm_add_pd(fjy1,ty);
 611             fjz1             = _mm_add_pd(fjz1,tz);
 612
 613             }
 614
 615             /**************************
 616              * CALCULATE INTERACTIONS *
 617              **************************/
 618
 619             if (gmx_mm_any_lt(rsq22,rcutoff2))
 620             {
 621
 622             /* REACTION-FIELD ELECTROSTATICS */
 623             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
 624             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
 625
 626             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 627
 628             /* Update potential sum for this i atom from the interaction with this j atom. */
 629             velec            = _mm_and_pd(velec,cutoff_mask);
 630             velecsum         = _mm_add_pd(velecsum,velec);
 631
 632             fscal            = felec;
 633
 634             fscal            = _mm_and_pd(fscal,cutoff_mask);
 635
 636             /* Calculate temporary vectorial force */
 637             tx               = _mm_mul_pd(fscal,dx22);
 638             ty               = _mm_mul_pd(fscal,dy22);
 639             tz               = _mm_mul_pd(fscal,dz22);
 640
 641             /* Update vectorial force */
 642             fix2             = _mm_add_pd(fix2,tx);
 643             fiy2             = _mm_add_pd(fiy2,ty);
 644             fiz2             = _mm_add_pd(fiz2,tz);
 645
 646             fjx2             = _mm_add_pd(fjx2,tx);
 647             fjy2             = _mm_add_pd(fjy2,ty);
 648             fjz2             = _mm_add_pd(fjz2,tz);
 649
 650             }
 651
 652             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 653
 654             /* Inner loop uses 360 flops */
 655         }
 656
 657         if(jidx<j_index_end)
 658         {
 659
 660             jnrA             = jjnr[jidx];
 661             j_coord_offsetA  = DIM*jnrA;
 662
 663             /* load j atom coordinates */
 664             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 665                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 666
 667             /* Calculate displacement vector */
 668             dx00             = _mm_sub_pd(ix0,jx0);
 669             dy00             = _mm_sub_pd(iy0,jy0);
 670             dz00             = _mm_sub_pd(iz0,jz0);
 671             dx01             = _mm_sub_pd(ix0,jx1);
 672             dy01             = _mm_sub_pd(iy0,jy1);
 673             dz01             = _mm_sub_pd(iz0,jz1);
 674             dx02             = _mm_sub_pd(ix0,jx2);
 675             dy02             = _mm_sub_pd(iy0,jy2);
 676             dz02             = _mm_sub_pd(iz0,jz2);
 677             dx10             = _mm_sub_pd(ix1,jx0);
 678             dy10             = _mm_sub_pd(iy1,jy0);
 679             dz10             = _mm_sub_pd(iz1,jz0);
 680             dx11             = _mm_sub_pd(ix1,jx1);
 681             dy11             = _mm_sub_pd(iy1,jy1);
 682             dz11             = _mm_sub_pd(iz1,jz1);
 683             dx12             = _mm_sub_pd(ix1,jx2);
 684             dy12             = _mm_sub_pd(iy1,jy2);
 685             dz12             = _mm_sub_pd(iz1,jz2);
 686             dx20             = _mm_sub_pd(ix2,jx0);
 687             dy20             = _mm_sub_pd(iy2,jy0);
 688             dz20             = _mm_sub_pd(iz2,jz0);
 689             dx21             = _mm_sub_pd(ix2,jx1);
 690             dy21             = _mm_sub_pd(iy2,jy1);
 691             dz21             = _mm_sub_pd(iz2,jz1);
 692             dx22             = _mm_sub_pd(ix2,jx2);
 693             dy22             = _mm_sub_pd(iy2,jy2);
 694             dz22             = _mm_sub_pd(iz2,jz2);
 695
 696             /* Calculate squared distance and things based on it */
 697             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 698             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 699             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 700             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 701             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 702             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 703             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 704             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 705             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 706
 707             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 708             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 709             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 710             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 711             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 712             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 713             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 714             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 715             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 716
 717             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 718             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 719             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 720             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 721             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 722             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 723             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 724             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 725             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 726
 727             fjx0             = _mm_setzero_pd();
 728             fjy0             = _mm_setzero_pd();
 729             fjz0             = _mm_setzero_pd();
 730             fjx1             = _mm_setzero_pd();
 731             fjy1             = _mm_setzero_pd();
 732             fjz1             = _mm_setzero_pd();
 733             fjx2             = _mm_setzero_pd();
 734             fjy2             = _mm_setzero_pd();
 735             fjz2             = _mm_setzero_pd();
 736
 737             /**************************
 738              * CALCULATE INTERACTIONS *
 739              **************************/
 740
 741             if (gmx_mm_any_lt(rsq00,rcutoff2))
 742             {
 743
 744             r00              = _mm_mul_pd(rsq00,rinv00);
 745
 746             /* Calculate table index by multiplying r with table scale and truncate to integer */
 747             rt               = _mm_mul_pd(r00,vftabscale);
 748             vfitab           = _mm_cvttpd_epi32(rt);
 749             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 750             vfitab           = _mm_slli_epi32(vfitab,3);
 751
 752             /* REACTION-FIELD ELECTROSTATICS */
 753             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_add_pd(rinv00,_mm_mul_pd(krf,rsq00)),crf));
 754             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
 755
 756             /* CUBIC SPLINE TABLE DISPERSION */
 757             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 758             F                = _mm_setzero_pd();
 759             GMX_MM_TRANSPOSE2_PD(Y,F);
 760             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 761             H                = _mm_setzero_pd();
 762             GMX_MM_TRANSPOSE2_PD(G,H);
 763             Heps             = _mm_mul_pd(vfeps,H);
 764             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 765             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 766             vvdw6            = _mm_mul_pd(c6_00,VV);
 767             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 768             fvdw6            = _mm_mul_pd(c6_00,FF);
 769
 770             /* CUBIC SPLINE TABLE REPULSION */
 771             vfitab           = _mm_add_epi32(vfitab,ifour);
 772             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 773             F                = _mm_setzero_pd();
 774             GMX_MM_TRANSPOSE2_PD(Y,F);
 775             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 776             H                = _mm_setzero_pd();
 777             GMX_MM_TRANSPOSE2_PD(G,H);
 778             Heps             = _mm_mul_pd(vfeps,H);
 779             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 780             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 781             vvdw12           = _mm_mul_pd(c12_00,VV);
 782             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 783             fvdw12           = _mm_mul_pd(c12_00,FF);
 784             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 785             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 786
 787             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 788
 789             /* Update potential sum for this i atom from the interaction with this j atom. */
 790             velec            = _mm_and_pd(velec,cutoff_mask);
 791             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 792             velecsum         = _mm_add_pd(velecsum,velec);
 793             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 794             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 795             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 796
 797             fscal            = _mm_add_pd(felec,fvdw);
 798
 799             fscal            = _mm_and_pd(fscal,cutoff_mask);
 800
 801             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 802
 803             /* Calculate temporary vectorial force */
 804             tx               = _mm_mul_pd(fscal,dx00);
 805             ty               = _mm_mul_pd(fscal,dy00);
 806             tz               = _mm_mul_pd(fscal,dz00);
 807
 808             /* Update vectorial force */
 809             fix0             = _mm_add_pd(fix0,tx);
 810             fiy0             = _mm_add_pd(fiy0,ty);
 811             fiz0             = _mm_add_pd(fiz0,tz);
 812
 813             fjx0             = _mm_add_pd(fjx0,tx);
 814             fjy0             = _mm_add_pd(fjy0,ty);
 815             fjz0             = _mm_add_pd(fjz0,tz);
 816
 817             }
 818
 819             /**************************
 820              * CALCULATE INTERACTIONS *
 821              **************************/
 822
 823             if (gmx_mm_any_lt(rsq01,rcutoff2))
 824             {
 825
 826             /* REACTION-FIELD ELECTROSTATICS */
 827             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_add_pd(rinv01,_mm_mul_pd(krf,rsq01)),crf));
 828             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
 829
 830             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
 831
 832             /* Update potential sum for this i atom from the interaction with this j atom. */
 833             velec            = _mm_and_pd(velec,cutoff_mask);
 834             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 835             velecsum         = _mm_add_pd(velecsum,velec);
 836
 837             fscal            = felec;
 838
 839             fscal            = _mm_and_pd(fscal,cutoff_mask);
 840
 841             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 842
 843             /* Calculate temporary vectorial force */
 844             tx               = _mm_mul_pd(fscal,dx01);
 845             ty               = _mm_mul_pd(fscal,dy01);
 846             tz               = _mm_mul_pd(fscal,dz01);
 847
 848             /* Update vectorial force */
 849             fix0             = _mm_add_pd(fix0,tx);
 850             fiy0             = _mm_add_pd(fiy0,ty);
 851             fiz0             = _mm_add_pd(fiz0,tz);
 852
 853             fjx1             = _mm_add_pd(fjx1,tx);
 854             fjy1             = _mm_add_pd(fjy1,ty);
 855             fjz1             = _mm_add_pd(fjz1,tz);
 856
 857             }
 858
 859             /**************************
 860              * CALCULATE INTERACTIONS *
 861              **************************/
 862
 863             if (gmx_mm_any_lt(rsq02,rcutoff2))
 864             {
 865
 866             /* REACTION-FIELD ELECTROSTATICS */
 867             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_add_pd(rinv02,_mm_mul_pd(krf,rsq02)),crf));
 868             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
 869
 870             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
 871
 872             /* Update potential sum for this i atom from the interaction with this j atom. */
 873             velec            = _mm_and_pd(velec,cutoff_mask);
 874             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 875             velecsum         = _mm_add_pd(velecsum,velec);
 876
 877             fscal            = felec;
 878
 879             fscal            = _mm_and_pd(fscal,cutoff_mask);
 880
 881             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 882
 883             /* Calculate temporary vectorial force */
 884             tx               = _mm_mul_pd(fscal,dx02);
 885             ty               = _mm_mul_pd(fscal,dy02);
 886             tz               = _mm_mul_pd(fscal,dz02);
 887
 888             /* Update vectorial force */
 889             fix0             = _mm_add_pd(fix0,tx);
 890             fiy0             = _mm_add_pd(fiy0,ty);
 891             fiz0             = _mm_add_pd(fiz0,tz);
 892
 893             fjx2             = _mm_add_pd(fjx2,tx);
 894             fjy2             = _mm_add_pd(fjy2,ty);
 895             fjz2             = _mm_add_pd(fjz2,tz);
 896
 897             }
 898
 899             /**************************
 900              * CALCULATE INTERACTIONS *
 901              **************************/
 902
 903             if (gmx_mm_any_lt(rsq10,rcutoff2))
 904             {
 905
 906             /* REACTION-FIELD ELECTROSTATICS */
 907             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_add_pd(rinv10,_mm_mul_pd(krf,rsq10)),crf));
 908             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
 909
 910             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
 911
 912             /* Update potential sum for this i atom from the interaction with this j atom. */
 913             velec            = _mm_and_pd(velec,cutoff_mask);
 914             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 915             velecsum         = _mm_add_pd(velecsum,velec);
 916
 917             fscal            = felec;
 918
 919             fscal            = _mm_and_pd(fscal,cutoff_mask);
 920
 921             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 922
 923             /* Calculate temporary vectorial force */
 924             tx               = _mm_mul_pd(fscal,dx10);
 925             ty               = _mm_mul_pd(fscal,dy10);
 926             tz               = _mm_mul_pd(fscal,dz10);
 927
 928             /* Update vectorial force */
 929             fix1             = _mm_add_pd(fix1,tx);
 930             fiy1             = _mm_add_pd(fiy1,ty);
 931             fiz1             = _mm_add_pd(fiz1,tz);
 932
 933             fjx0             = _mm_add_pd(fjx0,tx);
 934             fjy0             = _mm_add_pd(fjy0,ty);
 935             fjz0             = _mm_add_pd(fjz0,tz);
 936
 937             }
 938
 939             /**************************
 940              * CALCULATE INTERACTIONS *
 941              **************************/
 942
 943             if (gmx_mm_any_lt(rsq11,rcutoff2))
 944             {
 945
 946             /* REACTION-FIELD ELECTROSTATICS */
 947             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_add_pd(rinv11,_mm_mul_pd(krf,rsq11)),crf));
 948             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
 949
 950             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 951
 952             /* Update potential sum for this i atom from the interaction with this j atom. */
 953             velec            = _mm_and_pd(velec,cutoff_mask);
 954             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 955             velecsum         = _mm_add_pd(velecsum,velec);
 956
 957             fscal            = felec;
 958
 959             fscal            = _mm_and_pd(fscal,cutoff_mask);
 960
 961             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 962
 963             /* Calculate temporary vectorial force */
 964             tx               = _mm_mul_pd(fscal,dx11);
 965             ty               = _mm_mul_pd(fscal,dy11);
 966             tz               = _mm_mul_pd(fscal,dz11);
 967
 968             /* Update vectorial force */
 969             fix1             = _mm_add_pd(fix1,tx);
 970             fiy1             = _mm_add_pd(fiy1,ty);
 971             fiz1             = _mm_add_pd(fiz1,tz);
 972
 973             fjx1             = _mm_add_pd(fjx1,tx);
 974             fjy1             = _mm_add_pd(fjy1,ty);
 975             fjz1             = _mm_add_pd(fjz1,tz);
 976
 977             }
 978
 979             /**************************
 980              * CALCULATE INTERACTIONS *
 981              **************************/
 982
 983             if (gmx_mm_any_lt(rsq12,rcutoff2))
 984             {
 985
 986             /* REACTION-FIELD ELECTROSTATICS */
 987             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_add_pd(rinv12,_mm_mul_pd(krf,rsq12)),crf));
 988             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
 989
 990             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 991
 992             /* Update potential sum for this i atom from the interaction with this j atom. */
 993             velec            = _mm_and_pd(velec,cutoff_mask);
 994             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 995             velecsum         = _mm_add_pd(velecsum,velec);
 996
 997             fscal            = felec;
 998
 999             fscal            = _mm_and_pd(fscal,cutoff_mask);
1000
1001             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1002
1003             /* Calculate temporary vectorial force */
1004             tx               = _mm_mul_pd(fscal,dx12);
1005             ty               = _mm_mul_pd(fscal,dy12);
1006             tz               = _mm_mul_pd(fscal,dz12);
1007
1008             /* Update vectorial force */
1009             fix1             = _mm_add_pd(fix1,tx);
1010             fiy1             = _mm_add_pd(fiy1,ty);
1011             fiz1             = _mm_add_pd(fiz1,tz);
1012
1013             fjx2             = _mm_add_pd(fjx2,tx);
1014             fjy2             = _mm_add_pd(fjy2,ty);
1015             fjz2             = _mm_add_pd(fjz2,tz);
1016
1017             }
1018
1019             /**************************
1020              * CALCULATE INTERACTIONS *
1021              **************************/
1022
1023             if (gmx_mm_any_lt(rsq20,rcutoff2))
1024             {
1025
1026             /* REACTION-FIELD ELECTROSTATICS */
1027             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_add_pd(rinv20,_mm_mul_pd(krf,rsq20)),crf));
1028             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
1029
1030             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
1031
1032             /* Update potential sum for this i atom from the interaction with this j atom. */
1033             velec            = _mm_and_pd(velec,cutoff_mask);
1034             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1035             velecsum         = _mm_add_pd(velecsum,velec);
1036
1037             fscal            = felec;
1038
1039             fscal            = _mm_and_pd(fscal,cutoff_mask);
1040
1041             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1042
1043             /* Calculate temporary vectorial force */
1044             tx               = _mm_mul_pd(fscal,dx20);
1045             ty               = _mm_mul_pd(fscal,dy20);
1046             tz               = _mm_mul_pd(fscal,dz20);
1047
1048             /* Update vectorial force */
1049             fix2             = _mm_add_pd(fix2,tx);
1050             fiy2             = _mm_add_pd(fiy2,ty);
1051             fiz2             = _mm_add_pd(fiz2,tz);
1052
1053             fjx0             = _mm_add_pd(fjx0,tx);
1054             fjy0             = _mm_add_pd(fjy0,ty);
1055             fjz0             = _mm_add_pd(fjz0,tz);
1056
1057             }
1058
1059             /**************************
1060              * CALCULATE INTERACTIONS *
1061              **************************/
1062
1063             if (gmx_mm_any_lt(rsq21,rcutoff2))
1064             {
1065
1066             /* REACTION-FIELD ELECTROSTATICS */
1067             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_add_pd(rinv21,_mm_mul_pd(krf,rsq21)),crf));
1068             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1069
1070             cutoff_mask      = _mm_cmplt_pd(rsq21,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             /* Calculate temporary vectorial force */
1084             tx               = _mm_mul_pd(fscal,dx21);
1085             ty               = _mm_mul_pd(fscal,dy21);
1086             tz               = _mm_mul_pd(fscal,dz21);
1087
1088             /* Update vectorial force */
1089             fix2             = _mm_add_pd(fix2,tx);
1090             fiy2             = _mm_add_pd(fiy2,ty);
1091             fiz2             = _mm_add_pd(fiz2,tz);
1092
1093             fjx1             = _mm_add_pd(fjx1,tx);
1094             fjy1             = _mm_add_pd(fjy1,ty);
1095             fjz1             = _mm_add_pd(fjz1,tz);
1096
1097             }
1098
1099             /**************************
1100              * CALCULATE INTERACTIONS *
1101              **************************/
1102
1103             if (gmx_mm_any_lt(rsq22,rcutoff2))
1104             {
1105
1106             /* REACTION-FIELD ELECTROSTATICS */
1107             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_add_pd(rinv22,_mm_mul_pd(krf,rsq22)),crf));
1108             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1109
1110             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1111
1112             /* Update potential sum for this i atom from the interaction with this j atom. */
1113             velec            = _mm_and_pd(velec,cutoff_mask);
1114             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1115             velecsum         = _mm_add_pd(velecsum,velec);
1116
1117             fscal            = felec;
1118
1119             fscal            = _mm_and_pd(fscal,cutoff_mask);
1120
1121             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1122
1123             /* Calculate temporary vectorial force */
1124             tx               = _mm_mul_pd(fscal,dx22);
1125             ty               = _mm_mul_pd(fscal,dy22);
1126             tz               = _mm_mul_pd(fscal,dz22);
1127
1128             /* Update vectorial force */
1129             fix2             = _mm_add_pd(fix2,tx);
1130             fiy2             = _mm_add_pd(fiy2,ty);
1131             fiz2             = _mm_add_pd(fiz2,tz);
1132
1133             fjx2             = _mm_add_pd(fjx2,tx);
1134             fjy2             = _mm_add_pd(fjy2,ty);
1135             fjz2             = _mm_add_pd(fjz2,tz);
1136
1137             }
1138
1139             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1140
1141             /* Inner loop uses 360 flops */
1142         }
1143
1144         /* End of innermost loop */
1145
1146         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1147                                               f+i_coord_offset,fshift+i_shift_offset);
1148
1149         ggid                        = gid[iidx];
1150         /* Update potential energies */
1151         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1152         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
1153
1154         /* Increment number of inner iterations */
1155         inneriter                  += j_index_end - j_index_start;
1156
1157         /* Outer loop uses 20 flops */
1158     }
1159
1160     /* Increment number of outer iterations */
1161     outeriter        += nri;
1162
1163     /* Update outer/inner flops */
1164
1165     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*360);
1166 }
1167 /*
1168  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_F_sse4_1_double
1169  * Electrostatics interaction: ReactionField
1170  * VdW interaction:            CubicSplineTable
1171  * Geometry:                   Water3-Water3
1172  * Calculate force/pot:        Force
1173  */
1174 void
1175 nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_F_sse4_1_double
1176                     (t_nblist * gmx_restrict                nlist,
1177                      rvec * gmx_restrict                    xx,
1178                      rvec * gmx_restrict                    ff,
1179                      t_forcerec * gmx_restrict              fr,
1180                      t_mdatoms * gmx_restrict               mdatoms,
1181                      nb_kernel_data_t * gmx_restrict        kernel_data,
1182                      t_nrnb * gmx_restrict                  nrnb)
1183 {
1184     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1185      * just 0 for non-waters.
1186      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1187      * jnr indices corresponding to data put in the four positions in the SIMD register.
1188      */
1189     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1190     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1191     int              jnrA,jnrB;
1192     int              j_coord_offsetA,j_coord_offsetB;
1193     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1194     real             rcutoff_scalar;
1195     real             *shiftvec,*fshift,*x,*f;
1196     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1197     int              vdwioffset0;
1198     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1199     int              vdwioffset1;
1200     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1201     int              vdwioffset2;
1202     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1203     int              vdwjidx0A,vdwjidx0B;
1204     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1205     int              vdwjidx1A,vdwjidx1B;
1206     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1207     int              vdwjidx2A,vdwjidx2B;
1208     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1209     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1210     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1211     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1212     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1213     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1214     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1215     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1216     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1217     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1218     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1219     real             *charge;
1220     int              nvdwtype;
1221     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1222     int              *vdwtype;
1223     real             *vdwparam;
1224     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1225     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1226     __m128i          vfitab;
1227     __m128i          ifour       = _mm_set1_epi32(4);
1228     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
1229     real             *vftab;
1230     __m128d          dummy_mask,cutoff_mask;
1231     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1232     __m128d          one     = _mm_set1_pd(1.0);
1233     __m128d          two     = _mm_set1_pd(2.0);
1234     x                = xx[0];
1235     f                = ff[0];
1236
1237     nri              = nlist->nri;
1238     iinr             = nlist->iinr;
1239     jindex           = nlist->jindex;
1240     jjnr             = nlist->jjnr;
1241     shiftidx         = nlist->shift;
1242     gid              = nlist->gid;
1243     shiftvec         = fr->shift_vec[0];
1244     fshift           = fr->fshift[0];
1245     facel            = _mm_set1_pd(fr->epsfac);
1246     charge           = mdatoms->chargeA;
1247     krf              = _mm_set1_pd(fr->ic->k_rf);
1248     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1249     crf              = _mm_set1_pd(fr->ic->c_rf);
1250     nvdwtype         = fr->ntype;
1251     vdwparam         = fr->nbfp;
1252     vdwtype          = mdatoms->typeA;
1253
1254     vftab            = kernel_data->table_vdw->data;
1255     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
1256
1257     /* Setup water-specific parameters */
1258     inr              = nlist->iinr[0];
1259     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
1260     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1261     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1262     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1263
1264     jq0              = _mm_set1_pd(charge[inr+0]);
1265     jq1              = _mm_set1_pd(charge[inr+1]);
1266     jq2              = _mm_set1_pd(charge[inr+2]);
1267     vdwjidx0A        = 2*vdwtype[inr+0];
1268     qq00             = _mm_mul_pd(iq0,jq0);
1269     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1270     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1271     qq01             = _mm_mul_pd(iq0,jq1);
1272     qq02             = _mm_mul_pd(iq0,jq2);
1273     qq10             = _mm_mul_pd(iq1,jq0);
1274     qq11             = _mm_mul_pd(iq1,jq1);
1275     qq12             = _mm_mul_pd(iq1,jq2);
1276     qq20             = _mm_mul_pd(iq2,jq0);
1277     qq21             = _mm_mul_pd(iq2,jq1);
1278     qq22             = _mm_mul_pd(iq2,jq2);
1279
1280     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1281     rcutoff_scalar   = fr->rcoulomb;
1282     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1283     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1284
1285     /* Avoid stupid compiler warnings */
1286     jnrA = jnrB = 0;
1287     j_coord_offsetA = 0;
1288     j_coord_offsetB = 0;
1289
1290     outeriter        = 0;
1291     inneriter        = 0;
1292
1293     /* Start outer loop over neighborlists */
1294     for(iidx=0; iidx<nri; iidx++)
1295     {
1296         /* Load shift vector for this list */
1297         i_shift_offset   = DIM*shiftidx[iidx];
1298
1299         /* Load limits for loop over neighbors */
1300         j_index_start    = jindex[iidx];
1301         j_index_end      = jindex[iidx+1];
1302
1303         /* Get outer coordinate index */
1304         inr              = iinr[iidx];
1305         i_coord_offset   = DIM*inr;
1306
1307         /* Load i particle coords and add shift vector */
1308         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1309                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1310
1311         fix0             = _mm_setzero_pd();
1312         fiy0             = _mm_setzero_pd();
1313         fiz0             = _mm_setzero_pd();
1314         fix1             = _mm_setzero_pd();
1315         fiy1             = _mm_setzero_pd();
1316         fiz1             = _mm_setzero_pd();
1317         fix2             = _mm_setzero_pd();
1318         fiy2             = _mm_setzero_pd();
1319         fiz2             = _mm_setzero_pd();
1320
1321         /* Start inner kernel loop */
1322         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1323         {
1324
1325             /* Get j neighbor index, and coordinate index */
1326             jnrA             = jjnr[jidx];
1327             jnrB             = jjnr[jidx+1];
1328             j_coord_offsetA  = DIM*jnrA;
1329             j_coord_offsetB  = DIM*jnrB;
1330
1331             /* load j atom coordinates */
1332             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1333                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1334
1335             /* Calculate displacement vector */
1336             dx00             = _mm_sub_pd(ix0,jx0);
1337             dy00             = _mm_sub_pd(iy0,jy0);
1338             dz00             = _mm_sub_pd(iz0,jz0);
1339             dx01             = _mm_sub_pd(ix0,jx1);
1340             dy01             = _mm_sub_pd(iy0,jy1);
1341             dz01             = _mm_sub_pd(iz0,jz1);
1342             dx02             = _mm_sub_pd(ix0,jx2);
1343             dy02             = _mm_sub_pd(iy0,jy2);
1344             dz02             = _mm_sub_pd(iz0,jz2);
1345             dx10             = _mm_sub_pd(ix1,jx0);
1346             dy10             = _mm_sub_pd(iy1,jy0);
1347             dz10             = _mm_sub_pd(iz1,jz0);
1348             dx11             = _mm_sub_pd(ix1,jx1);
1349             dy11             = _mm_sub_pd(iy1,jy1);
1350             dz11             = _mm_sub_pd(iz1,jz1);
1351             dx12             = _mm_sub_pd(ix1,jx2);
1352             dy12             = _mm_sub_pd(iy1,jy2);
1353             dz12             = _mm_sub_pd(iz1,jz2);
1354             dx20             = _mm_sub_pd(ix2,jx0);
1355             dy20             = _mm_sub_pd(iy2,jy0);
1356             dz20             = _mm_sub_pd(iz2,jz0);
1357             dx21             = _mm_sub_pd(ix2,jx1);
1358             dy21             = _mm_sub_pd(iy2,jy1);
1359             dz21             = _mm_sub_pd(iz2,jz1);
1360             dx22             = _mm_sub_pd(ix2,jx2);
1361             dy22             = _mm_sub_pd(iy2,jy2);
1362             dz22             = _mm_sub_pd(iz2,jz2);
1363
1364             /* Calculate squared distance and things based on it */
1365             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1366             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1367             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1368             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1369             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1370             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1371             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1372             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1373             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1374
1375             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1376             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1377             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1378             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1379             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1380             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1381             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1382             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1383             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1384
1385             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1386             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1387             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1388             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1389             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1390             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1391             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1392             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1393             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1394
1395             fjx0             = _mm_setzero_pd();
1396             fjy0             = _mm_setzero_pd();
1397             fjz0             = _mm_setzero_pd();
1398             fjx1             = _mm_setzero_pd();
1399             fjy1             = _mm_setzero_pd();
1400             fjz1             = _mm_setzero_pd();
1401             fjx2             = _mm_setzero_pd();
1402             fjy2             = _mm_setzero_pd();
1403             fjz2             = _mm_setzero_pd();
1404
1405             /**************************
1406              * CALCULATE INTERACTIONS *
1407              **************************/
1408
1409             if (gmx_mm_any_lt(rsq00,rcutoff2))
1410             {
1411
1412             r00              = _mm_mul_pd(rsq00,rinv00);
1413
1414             /* Calculate table index by multiplying r with table scale and truncate to integer */
1415             rt               = _mm_mul_pd(r00,vftabscale);
1416             vfitab           = _mm_cvttpd_epi32(rt);
1417             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
1418             vfitab           = _mm_slli_epi32(vfitab,3);
1419
1420             /* REACTION-FIELD ELECTROSTATICS */
1421             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
1422
1423             /* CUBIC SPLINE TABLE DISPERSION */
1424             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1425             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
1426             GMX_MM_TRANSPOSE2_PD(Y,F);
1427             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1428             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
1429             GMX_MM_TRANSPOSE2_PD(G,H);
1430             Heps             = _mm_mul_pd(vfeps,H);
1431             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1432             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1433             fvdw6            = _mm_mul_pd(c6_00,FF);
1434
1435             /* CUBIC SPLINE TABLE REPULSION */
1436             vfitab           = _mm_add_epi32(vfitab,ifour);
1437             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1438             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
1439             GMX_MM_TRANSPOSE2_PD(Y,F);
1440             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1441             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
1442             GMX_MM_TRANSPOSE2_PD(G,H);
1443             Heps             = _mm_mul_pd(vfeps,H);
1444             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1445             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1446             fvdw12           = _mm_mul_pd(c12_00,FF);
1447             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1448
1449             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1450
1451             fscal            = _mm_add_pd(felec,fvdw);
1452
1453             fscal            = _mm_and_pd(fscal,cutoff_mask);
1454
1455             /* Calculate temporary vectorial force */
1456             tx               = _mm_mul_pd(fscal,dx00);
1457             ty               = _mm_mul_pd(fscal,dy00);
1458             tz               = _mm_mul_pd(fscal,dz00);
1459
1460             /* Update vectorial force */
1461             fix0             = _mm_add_pd(fix0,tx);
1462             fiy0             = _mm_add_pd(fiy0,ty);
1463             fiz0             = _mm_add_pd(fiz0,tz);
1464
1465             fjx0             = _mm_add_pd(fjx0,tx);
1466             fjy0             = _mm_add_pd(fjy0,ty);
1467             fjz0             = _mm_add_pd(fjz0,tz);
1468
1469             }
1470
1471             /**************************
1472              * CALCULATE INTERACTIONS *
1473              **************************/
1474
1475             if (gmx_mm_any_lt(rsq01,rcutoff2))
1476             {
1477
1478             /* REACTION-FIELD ELECTROSTATICS */
1479             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
1480
1481             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
1482
1483             fscal            = felec;
1484
1485             fscal            = _mm_and_pd(fscal,cutoff_mask);
1486
1487             /* Calculate temporary vectorial force */
1488             tx               = _mm_mul_pd(fscal,dx01);
1489             ty               = _mm_mul_pd(fscal,dy01);
1490             tz               = _mm_mul_pd(fscal,dz01);
1491
1492             /* Update vectorial force */
1493             fix0             = _mm_add_pd(fix0,tx);
1494             fiy0             = _mm_add_pd(fiy0,ty);
1495             fiz0             = _mm_add_pd(fiz0,tz);
1496
1497             fjx1             = _mm_add_pd(fjx1,tx);
1498             fjy1             = _mm_add_pd(fjy1,ty);
1499             fjz1             = _mm_add_pd(fjz1,tz);
1500
1501             }
1502
1503             /**************************
1504              * CALCULATE INTERACTIONS *
1505              **************************/
1506
1507             if (gmx_mm_any_lt(rsq02,rcutoff2))
1508             {
1509
1510             /* REACTION-FIELD ELECTROSTATICS */
1511             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
1512
1513             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
1514
1515             fscal            = felec;
1516
1517             fscal            = _mm_and_pd(fscal,cutoff_mask);
1518
1519             /* Calculate temporary vectorial force */
1520             tx               = _mm_mul_pd(fscal,dx02);
1521             ty               = _mm_mul_pd(fscal,dy02);
1522             tz               = _mm_mul_pd(fscal,dz02);
1523
1524             /* Update vectorial force */
1525             fix0             = _mm_add_pd(fix0,tx);
1526             fiy0             = _mm_add_pd(fiy0,ty);
1527             fiz0             = _mm_add_pd(fiz0,tz);
1528
1529             fjx2             = _mm_add_pd(fjx2,tx);
1530             fjy2             = _mm_add_pd(fjy2,ty);
1531             fjz2             = _mm_add_pd(fjz2,tz);
1532
1533             }
1534
1535             /**************************
1536              * CALCULATE INTERACTIONS *
1537              **************************/
1538
1539             if (gmx_mm_any_lt(rsq10,rcutoff2))
1540             {
1541
1542             /* REACTION-FIELD ELECTROSTATICS */
1543             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
1544
1545             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
1546
1547             fscal            = felec;
1548
1549             fscal            = _mm_and_pd(fscal,cutoff_mask);
1550
1551             /* Calculate temporary vectorial force */
1552             tx               = _mm_mul_pd(fscal,dx10);
1553             ty               = _mm_mul_pd(fscal,dy10);
1554             tz               = _mm_mul_pd(fscal,dz10);
1555
1556             /* Update vectorial force */
1557             fix1             = _mm_add_pd(fix1,tx);
1558             fiy1             = _mm_add_pd(fiy1,ty);
1559             fiz1             = _mm_add_pd(fiz1,tz);
1560
1561             fjx0             = _mm_add_pd(fjx0,tx);
1562             fjy0             = _mm_add_pd(fjy0,ty);
1563             fjz0             = _mm_add_pd(fjz0,tz);
1564
1565             }
1566
1567             /**************************
1568              * CALCULATE INTERACTIONS *
1569              **************************/
1570
1571             if (gmx_mm_any_lt(rsq11,rcutoff2))
1572             {
1573
1574             /* REACTION-FIELD ELECTROSTATICS */
1575             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1576
1577             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1578
1579             fscal            = felec;
1580
1581             fscal            = _mm_and_pd(fscal,cutoff_mask);
1582
1583             /* Calculate temporary vectorial force */
1584             tx               = _mm_mul_pd(fscal,dx11);
1585             ty               = _mm_mul_pd(fscal,dy11);
1586             tz               = _mm_mul_pd(fscal,dz11);
1587
1588             /* Update vectorial force */
1589             fix1             = _mm_add_pd(fix1,tx);
1590             fiy1             = _mm_add_pd(fiy1,ty);
1591             fiz1             = _mm_add_pd(fiz1,tz);
1592
1593             fjx1             = _mm_add_pd(fjx1,tx);
1594             fjy1             = _mm_add_pd(fjy1,ty);
1595             fjz1             = _mm_add_pd(fjz1,tz);
1596
1597             }
1598
1599             /**************************
1600              * CALCULATE INTERACTIONS *
1601              **************************/
1602
1603             if (gmx_mm_any_lt(rsq12,rcutoff2))
1604             {
1605
1606             /* REACTION-FIELD ELECTROSTATICS */
1607             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
1608
1609             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1610
1611             fscal            = felec;
1612
1613             fscal            = _mm_and_pd(fscal,cutoff_mask);
1614
1615             /* Calculate temporary vectorial force */
1616             tx               = _mm_mul_pd(fscal,dx12);
1617             ty               = _mm_mul_pd(fscal,dy12);
1618             tz               = _mm_mul_pd(fscal,dz12);
1619
1620             /* Update vectorial force */
1621             fix1             = _mm_add_pd(fix1,tx);
1622             fiy1             = _mm_add_pd(fiy1,ty);
1623             fiz1             = _mm_add_pd(fiz1,tz);
1624
1625             fjx2             = _mm_add_pd(fjx2,tx);
1626             fjy2             = _mm_add_pd(fjy2,ty);
1627             fjz2             = _mm_add_pd(fjz2,tz);
1628
1629             }
1630
1631             /**************************
1632              * CALCULATE INTERACTIONS *
1633              **************************/
1634
1635             if (gmx_mm_any_lt(rsq20,rcutoff2))
1636             {
1637
1638             /* REACTION-FIELD ELECTROSTATICS */
1639             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
1640
1641             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
1642
1643             fscal            = felec;
1644
1645             fscal            = _mm_and_pd(fscal,cutoff_mask);
1646
1647             /* Calculate temporary vectorial force */
1648             tx               = _mm_mul_pd(fscal,dx20);
1649             ty               = _mm_mul_pd(fscal,dy20);
1650             tz               = _mm_mul_pd(fscal,dz20);
1651
1652             /* Update vectorial force */
1653             fix2             = _mm_add_pd(fix2,tx);
1654             fiy2             = _mm_add_pd(fiy2,ty);
1655             fiz2             = _mm_add_pd(fiz2,tz);
1656
1657             fjx0             = _mm_add_pd(fjx0,tx);
1658             fjy0             = _mm_add_pd(fjy0,ty);
1659             fjz0             = _mm_add_pd(fjz0,tz);
1660
1661             }
1662
1663             /**************************
1664              * CALCULATE INTERACTIONS *
1665              **************************/
1666
1667             if (gmx_mm_any_lt(rsq21,rcutoff2))
1668             {
1669
1670             /* REACTION-FIELD ELECTROSTATICS */
1671             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
1672
1673             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1674
1675             fscal            = felec;
1676
1677             fscal            = _mm_and_pd(fscal,cutoff_mask);
1678
1679             /* Calculate temporary vectorial force */
1680             tx               = _mm_mul_pd(fscal,dx21);
1681             ty               = _mm_mul_pd(fscal,dy21);
1682             tz               = _mm_mul_pd(fscal,dz21);
1683
1684             /* Update vectorial force */
1685             fix2             = _mm_add_pd(fix2,tx);
1686             fiy2             = _mm_add_pd(fiy2,ty);
1687             fiz2             = _mm_add_pd(fiz2,tz);
1688
1689             fjx1             = _mm_add_pd(fjx1,tx);
1690             fjy1             = _mm_add_pd(fjy1,ty);
1691             fjz1             = _mm_add_pd(fjz1,tz);
1692
1693             }
1694
1695             /**************************
1696              * CALCULATE INTERACTIONS *
1697              **************************/
1698
1699             if (gmx_mm_any_lt(rsq22,rcutoff2))
1700             {
1701
1702             /* REACTION-FIELD ELECTROSTATICS */
1703             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
1704
1705             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1706
1707             fscal            = felec;
1708
1709             fscal            = _mm_and_pd(fscal,cutoff_mask);
1710
1711             /* Calculate temporary vectorial force */
1712             tx               = _mm_mul_pd(fscal,dx22);
1713             ty               = _mm_mul_pd(fscal,dy22);
1714             tz               = _mm_mul_pd(fscal,dz22);
1715
1716             /* Update vectorial force */
1717             fix2             = _mm_add_pd(fix2,tx);
1718             fiy2             = _mm_add_pd(fiy2,ty);
1719             fiz2             = _mm_add_pd(fiz2,tz);
1720
1721             fjx2             = _mm_add_pd(fjx2,tx);
1722             fjy2             = _mm_add_pd(fjy2,ty);
1723             fjz2             = _mm_add_pd(fjz2,tz);
1724
1725             }
1726
1727             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1728
1729             /* Inner loop uses 297 flops */
1730         }
1731
1732         if(jidx<j_index_end)
1733         {
1734
1735             jnrA             = jjnr[jidx];
1736             j_coord_offsetA  = DIM*jnrA;
1737
1738             /* load j atom coordinates */
1739             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1740                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1741
1742             /* Calculate displacement vector */
1743             dx00             = _mm_sub_pd(ix0,jx0);
1744             dy00             = _mm_sub_pd(iy0,jy0);
1745             dz00             = _mm_sub_pd(iz0,jz0);
1746             dx01             = _mm_sub_pd(ix0,jx1);
1747             dy01             = _mm_sub_pd(iy0,jy1);
1748             dz01             = _mm_sub_pd(iz0,jz1);
1749             dx02             = _mm_sub_pd(ix0,jx2);
1750             dy02             = _mm_sub_pd(iy0,jy2);
1751             dz02             = _mm_sub_pd(iz0,jz2);
1752             dx10             = _mm_sub_pd(ix1,jx0);
1753             dy10             = _mm_sub_pd(iy1,jy0);
1754             dz10             = _mm_sub_pd(iz1,jz0);
1755             dx11             = _mm_sub_pd(ix1,jx1);
1756             dy11             = _mm_sub_pd(iy1,jy1);
1757             dz11             = _mm_sub_pd(iz1,jz1);
1758             dx12             = _mm_sub_pd(ix1,jx2);
1759             dy12             = _mm_sub_pd(iy1,jy2);
1760             dz12             = _mm_sub_pd(iz1,jz2);
1761             dx20             = _mm_sub_pd(ix2,jx0);
1762             dy20             = _mm_sub_pd(iy2,jy0);
1763             dz20             = _mm_sub_pd(iz2,jz0);
1764             dx21             = _mm_sub_pd(ix2,jx1);
1765             dy21             = _mm_sub_pd(iy2,jy1);
1766             dz21             = _mm_sub_pd(iz2,jz1);
1767             dx22             = _mm_sub_pd(ix2,jx2);
1768             dy22             = _mm_sub_pd(iy2,jy2);
1769             dz22             = _mm_sub_pd(iz2,jz2);
1770
1771             /* Calculate squared distance and things based on it */
1772             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1773             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1774             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1775             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1776             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1777             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1778             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1779             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1780             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1781
1782             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1783             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1784             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1785             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1786             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1787             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1788             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1789             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1790             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1791
1792             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1793             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1794             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1795             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1796             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1797             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1798             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1799             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1800             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1801
1802             fjx0             = _mm_setzero_pd();
1803             fjy0             = _mm_setzero_pd();
1804             fjz0             = _mm_setzero_pd();
1805             fjx1             = _mm_setzero_pd();
1806             fjy1             = _mm_setzero_pd();
1807             fjz1             = _mm_setzero_pd();
1808             fjx2             = _mm_setzero_pd();
1809             fjy2             = _mm_setzero_pd();
1810             fjz2             = _mm_setzero_pd();
1811
1812             /**************************
1813              * CALCULATE INTERACTIONS *
1814              **************************/
1815
1816             if (gmx_mm_any_lt(rsq00,rcutoff2))
1817             {
1818
1819             r00              = _mm_mul_pd(rsq00,rinv00);
1820
1821             /* Calculate table index by multiplying r with table scale and truncate to integer */
1822             rt               = _mm_mul_pd(r00,vftabscale);
1823             vfitab           = _mm_cvttpd_epi32(rt);
1824             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
1825             vfitab           = _mm_slli_epi32(vfitab,3);
1826
1827             /* REACTION-FIELD ELECTROSTATICS */
1828             felec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_mul_pd(rinv00,rinvsq00),krf2));
1829
1830             /* CUBIC SPLINE TABLE DISPERSION */
1831             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1832             F                = _mm_setzero_pd();
1833             GMX_MM_TRANSPOSE2_PD(Y,F);
1834             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1835             H                = _mm_setzero_pd();
1836             GMX_MM_TRANSPOSE2_PD(G,H);
1837             Heps             = _mm_mul_pd(vfeps,H);
1838             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1839             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1840             fvdw6            = _mm_mul_pd(c6_00,FF);
1841
1842             /* CUBIC SPLINE TABLE REPULSION */
1843             vfitab           = _mm_add_epi32(vfitab,ifour);
1844             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1845             F                = _mm_setzero_pd();
1846             GMX_MM_TRANSPOSE2_PD(Y,F);
1847             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1848             H                = _mm_setzero_pd();
1849             GMX_MM_TRANSPOSE2_PD(G,H);
1850             Heps             = _mm_mul_pd(vfeps,H);
1851             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1852             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1853             fvdw12           = _mm_mul_pd(c12_00,FF);
1854             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1855
1856             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1857
1858             fscal            = _mm_add_pd(felec,fvdw);
1859
1860             fscal            = _mm_and_pd(fscal,cutoff_mask);
1861
1862             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1863
1864             /* Calculate temporary vectorial force */
1865             tx               = _mm_mul_pd(fscal,dx00);
1866             ty               = _mm_mul_pd(fscal,dy00);
1867             tz               = _mm_mul_pd(fscal,dz00);
1868
1869             /* Update vectorial force */
1870             fix0             = _mm_add_pd(fix0,tx);
1871             fiy0             = _mm_add_pd(fiy0,ty);
1872             fiz0             = _mm_add_pd(fiz0,tz);
1873
1874             fjx0             = _mm_add_pd(fjx0,tx);
1875             fjy0             = _mm_add_pd(fjy0,ty);
1876             fjz0             = _mm_add_pd(fjz0,tz);
1877
1878             }
1879
1880             /**************************
1881              * CALCULATE INTERACTIONS *
1882              **************************/
1883
1884             if (gmx_mm_any_lt(rsq01,rcutoff2))
1885             {
1886
1887             /* REACTION-FIELD ELECTROSTATICS */
1888             felec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_mul_pd(rinv01,rinvsq01),krf2));
1889
1890             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
1891
1892             fscal            = felec;
1893
1894             fscal            = _mm_and_pd(fscal,cutoff_mask);
1895
1896             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1897
1898             /* Calculate temporary vectorial force */
1899             tx               = _mm_mul_pd(fscal,dx01);
1900             ty               = _mm_mul_pd(fscal,dy01);
1901             tz               = _mm_mul_pd(fscal,dz01);
1902
1903             /* Update vectorial force */
1904             fix0             = _mm_add_pd(fix0,tx);
1905             fiy0             = _mm_add_pd(fiy0,ty);
1906             fiz0             = _mm_add_pd(fiz0,tz);
1907
1908             fjx1             = _mm_add_pd(fjx1,tx);
1909             fjy1             = _mm_add_pd(fjy1,ty);
1910             fjz1             = _mm_add_pd(fjz1,tz);
1911
1912             }
1913
1914             /**************************
1915              * CALCULATE INTERACTIONS *
1916              **************************/
1917
1918             if (gmx_mm_any_lt(rsq02,rcutoff2))
1919             {
1920
1921             /* REACTION-FIELD ELECTROSTATICS */
1922             felec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_mul_pd(rinv02,rinvsq02),krf2));
1923
1924             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
1925
1926             fscal            = felec;
1927
1928             fscal            = _mm_and_pd(fscal,cutoff_mask);
1929
1930             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1931
1932             /* Calculate temporary vectorial force */
1933             tx               = _mm_mul_pd(fscal,dx02);
1934             ty               = _mm_mul_pd(fscal,dy02);
1935             tz               = _mm_mul_pd(fscal,dz02);
1936
1937             /* Update vectorial force */
1938             fix0             = _mm_add_pd(fix0,tx);
1939             fiy0             = _mm_add_pd(fiy0,ty);
1940             fiz0             = _mm_add_pd(fiz0,tz);
1941
1942             fjx2             = _mm_add_pd(fjx2,tx);
1943             fjy2             = _mm_add_pd(fjy2,ty);
1944             fjz2             = _mm_add_pd(fjz2,tz);
1945
1946             }
1947
1948             /**************************
1949              * CALCULATE INTERACTIONS *
1950              **************************/
1951
1952             if (gmx_mm_any_lt(rsq10,rcutoff2))
1953             {
1954
1955             /* REACTION-FIELD ELECTROSTATICS */
1956             felec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_mul_pd(rinv10,rinvsq10),krf2));
1957
1958             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
1959
1960             fscal            = felec;
1961
1962             fscal            = _mm_and_pd(fscal,cutoff_mask);
1963
1964             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1965
1966             /* Calculate temporary vectorial force */
1967             tx               = _mm_mul_pd(fscal,dx10);
1968             ty               = _mm_mul_pd(fscal,dy10);
1969             tz               = _mm_mul_pd(fscal,dz10);
1970
1971             /* Update vectorial force */
1972             fix1             = _mm_add_pd(fix1,tx);
1973             fiy1             = _mm_add_pd(fiy1,ty);
1974             fiz1             = _mm_add_pd(fiz1,tz);
1975
1976             fjx0             = _mm_add_pd(fjx0,tx);
1977             fjy0             = _mm_add_pd(fjy0,ty);
1978             fjz0             = _mm_add_pd(fjz0,tz);
1979
1980             }
1981
1982             /**************************
1983              * CALCULATE INTERACTIONS *
1984              **************************/
1985
1986             if (gmx_mm_any_lt(rsq11,rcutoff2))
1987             {
1988
1989             /* REACTION-FIELD ELECTROSTATICS */
1990             felec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_mul_pd(rinv11,rinvsq11),krf2));
1991
1992             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1993
1994             fscal            = felec;
1995
1996             fscal            = _mm_and_pd(fscal,cutoff_mask);
1997
1998             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1999
2000             /* Calculate temporary vectorial force */
2001             tx               = _mm_mul_pd(fscal,dx11);
2002             ty               = _mm_mul_pd(fscal,dy11);
2003             tz               = _mm_mul_pd(fscal,dz11);
2004
2005             /* Update vectorial force */
2006             fix1             = _mm_add_pd(fix1,tx);
2007             fiy1             = _mm_add_pd(fiy1,ty);
2008             fiz1             = _mm_add_pd(fiz1,tz);
2009
2010             fjx1             = _mm_add_pd(fjx1,tx);
2011             fjy1             = _mm_add_pd(fjy1,ty);
2012             fjz1             = _mm_add_pd(fjz1,tz);
2013
2014             }
2015
2016             /**************************
2017              * CALCULATE INTERACTIONS *
2018              **************************/
2019
2020             if (gmx_mm_any_lt(rsq12,rcutoff2))
2021             {
2022
2023             /* REACTION-FIELD ELECTROSTATICS */
2024             felec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_mul_pd(rinv12,rinvsq12),krf2));
2025
2026             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
2027
2028             fscal            = felec;
2029
2030             fscal            = _mm_and_pd(fscal,cutoff_mask);
2031
2032             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2033
2034             /* Calculate temporary vectorial force */
2035             tx               = _mm_mul_pd(fscal,dx12);
2036             ty               = _mm_mul_pd(fscal,dy12);
2037             tz               = _mm_mul_pd(fscal,dz12);
2038
2039             /* Update vectorial force */
2040             fix1             = _mm_add_pd(fix1,tx);
2041             fiy1             = _mm_add_pd(fiy1,ty);
2042             fiz1             = _mm_add_pd(fiz1,tz);
2043
2044             fjx2             = _mm_add_pd(fjx2,tx);
2045             fjy2             = _mm_add_pd(fjy2,ty);
2046             fjz2             = _mm_add_pd(fjz2,tz);
2047
2048             }
2049
2050             /**************************
2051              * CALCULATE INTERACTIONS *
2052              **************************/
2053
2054             if (gmx_mm_any_lt(rsq20,rcutoff2))
2055             {
2056
2057             /* REACTION-FIELD ELECTROSTATICS */
2058             felec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_mul_pd(rinv20,rinvsq20),krf2));
2059
2060             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
2061
2062             fscal            = felec;
2063
2064             fscal            = _mm_and_pd(fscal,cutoff_mask);
2065
2066             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2067
2068             /* Calculate temporary vectorial force */
2069             tx               = _mm_mul_pd(fscal,dx20);
2070             ty               = _mm_mul_pd(fscal,dy20);
2071             tz               = _mm_mul_pd(fscal,dz20);
2072
2073             /* Update vectorial force */
2074             fix2             = _mm_add_pd(fix2,tx);
2075             fiy2             = _mm_add_pd(fiy2,ty);
2076             fiz2             = _mm_add_pd(fiz2,tz);
2077
2078             fjx0             = _mm_add_pd(fjx0,tx);
2079             fjy0             = _mm_add_pd(fjy0,ty);
2080             fjz0             = _mm_add_pd(fjz0,tz);
2081
2082             }
2083
2084             /**************************
2085              * CALCULATE INTERACTIONS *
2086              **************************/
2087
2088             if (gmx_mm_any_lt(rsq21,rcutoff2))
2089             {
2090
2091             /* REACTION-FIELD ELECTROSTATICS */
2092             felec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_mul_pd(rinv21,rinvsq21),krf2));
2093
2094             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
2095
2096             fscal            = felec;
2097
2098             fscal            = _mm_and_pd(fscal,cutoff_mask);
2099
2100             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2101
2102             /* Calculate temporary vectorial force */
2103             tx               = _mm_mul_pd(fscal,dx21);
2104             ty               = _mm_mul_pd(fscal,dy21);
2105             tz               = _mm_mul_pd(fscal,dz21);
2106
2107             /* Update vectorial force */
2108             fix2             = _mm_add_pd(fix2,tx);
2109             fiy2             = _mm_add_pd(fiy2,ty);
2110             fiz2             = _mm_add_pd(fiz2,tz);
2111
2112             fjx1             = _mm_add_pd(fjx1,tx);
2113             fjy1             = _mm_add_pd(fjy1,ty);
2114             fjz1             = _mm_add_pd(fjz1,tz);
2115
2116             }
2117
2118             /**************************
2119              * CALCULATE INTERACTIONS *
2120              **************************/
2121
2122             if (gmx_mm_any_lt(rsq22,rcutoff2))
2123             {
2124
2125             /* REACTION-FIELD ELECTROSTATICS */
2126             felec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_mul_pd(rinv22,rinvsq22),krf2));
2127
2128             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
2129
2130             fscal            = felec;
2131
2132             fscal            = _mm_and_pd(fscal,cutoff_mask);
2133
2134             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2135
2136             /* Calculate temporary vectorial force */
2137             tx               = _mm_mul_pd(fscal,dx22);
2138             ty               = _mm_mul_pd(fscal,dy22);
2139             tz               = _mm_mul_pd(fscal,dz22);
2140
2141             /* Update vectorial force */
2142             fix2             = _mm_add_pd(fix2,tx);
2143             fiy2             = _mm_add_pd(fiy2,ty);
2144             fiz2             = _mm_add_pd(fiz2,tz);
2145
2146             fjx2             = _mm_add_pd(fjx2,tx);
2147             fjy2             = _mm_add_pd(fjy2,ty);
2148             fjz2             = _mm_add_pd(fjz2,tz);
2149
2150             }
2151
2152             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2153
2154             /* Inner loop uses 297 flops */
2155         }
2156
2157         /* End of innermost loop */
2158
2159         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
2160                                               f+i_coord_offset,fshift+i_shift_offset);
2161
2162         /* Increment number of inner iterations */
2163         inneriter                  += j_index_end - j_index_start;
2164
2165         /* Outer loop uses 18 flops */
2166     }
2167
2168     /* Increment number of outer iterations */
2169     outeriter        += nri;
2170
2171     /* Update outer/inner flops */
2172
2173     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*297);
2174 }