src/gmxlib/nonbonded/nb_kernel_sse4_1_double/nb_kernel_ElecEwSh_VdwNone_GeomW4W4_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_ElecEwSh_VdwNone_GeomW4W4_VF_sse4_1_double
  38  * Electrostatics interaction: Ewald
  39  * VdW interaction:            None
  40  * Geometry:                   Water4-Water4
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecEwSh_VdwNone_GeomW4W4_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              vdwioffset1;
  67     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  68     int              vdwioffset2;
  69     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  70     int              vdwioffset3;
  71     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  72     int              vdwjidx1A,vdwjidx1B;
  73     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  74     int              vdwjidx2A,vdwjidx2B;
  75     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  76     int              vdwjidx3A,vdwjidx3B;
  77     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  78     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  79     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  80     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  81     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  82     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  83     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  84     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  85     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
  86     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
  87     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  88     real             *charge;
  89     __m128i          ewitab;
  90     __m128d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
  91     real             *ewtab;
  92     __m128d          dummy_mask,cutoff_mask;
  93     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
  94     __m128d          one     = _mm_set1_pd(1.0);
  95     __m128d          two     = _mm_set1_pd(2.0);
  96     x                = xx[0];
  97     f                = ff[0];
  98
  99     nri              = nlist->nri;
 100     iinr             = nlist->iinr;
 101     jindex           = nlist->jindex;
 102     jjnr             = nlist->jjnr;
 103     shiftidx         = nlist->shift;
 104     gid              = nlist->gid;
 105     shiftvec         = fr->shift_vec[0];
 106     fshift           = fr->fshift[0];
 107     facel            = _mm_set1_pd(fr->epsfac);
 108     charge           = mdatoms->chargeA;
 109
 110     sh_ewald         = _mm_set1_pd(fr->ic->sh_ewald);
 111     ewtab            = fr->ic->tabq_coul_FDV0;
 112     ewtabscale       = _mm_set1_pd(fr->ic->tabq_scale);
 113     ewtabhalfspace   = _mm_set1_pd(0.5/fr->ic->tabq_scale);
 114
 115     /* Setup water-specific parameters */
 116     inr              = nlist->iinr[0];
 117     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 118     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 119     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 120
 121     jq1              = _mm_set1_pd(charge[inr+1]);
 122     jq2              = _mm_set1_pd(charge[inr+2]);
 123     jq3              = _mm_set1_pd(charge[inr+3]);
 124     qq11             = _mm_mul_pd(iq1,jq1);
 125     qq12             = _mm_mul_pd(iq1,jq2);
 126     qq13             = _mm_mul_pd(iq1,jq3);
 127     qq21             = _mm_mul_pd(iq2,jq1);
 128     qq22             = _mm_mul_pd(iq2,jq2);
 129     qq23             = _mm_mul_pd(iq2,jq3);
 130     qq31             = _mm_mul_pd(iq3,jq1);
 131     qq32             = _mm_mul_pd(iq3,jq2);
 132     qq33             = _mm_mul_pd(iq3,jq3);
 133
 134     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 135     rcutoff_scalar   = fr->rcoulomb;
 136     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 137     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 138
 139     /* Avoid stupid compiler warnings */
 140     jnrA = jnrB = 0;
 141     j_coord_offsetA = 0;
 142     j_coord_offsetB = 0;
 143
 144     outeriter        = 0;
 145     inneriter        = 0;
 146
 147     /* Start outer loop over neighborlists */
 148     for(iidx=0; iidx<nri; iidx++)
 149     {
 150         /* Load shift vector for this list */
 151         i_shift_offset   = DIM*shiftidx[iidx];
 152
 153         /* Load limits for loop over neighbors */
 154         j_index_start    = jindex[iidx];
 155         j_index_end      = jindex[iidx+1];
 156
 157         /* Get outer coordinate index */
 158         inr              = iinr[iidx];
 159         i_coord_offset   = DIM*inr;
 160
 161         /* Load i particle coords and add shift vector */
 162         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
 163                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 164
 165         fix1             = _mm_setzero_pd();
 166         fiy1             = _mm_setzero_pd();
 167         fiz1             = _mm_setzero_pd();
 168         fix2             = _mm_setzero_pd();
 169         fiy2             = _mm_setzero_pd();
 170         fiz2             = _mm_setzero_pd();
 171         fix3             = _mm_setzero_pd();
 172         fiy3             = _mm_setzero_pd();
 173         fiz3             = _mm_setzero_pd();
 174
 175         /* Reset potential sums */
 176         velecsum         = _mm_setzero_pd();
 177
 178         /* Start inner kernel loop */
 179         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 180         {
 181
 182             /* Get j neighbor index, and coordinate index */
 183             jnrA             = jjnr[jidx];
 184             jnrB             = jjnr[jidx+1];
 185             j_coord_offsetA  = DIM*jnrA;
 186             j_coord_offsetB  = DIM*jnrB;
 187
 188             /* load j atom coordinates */
 189             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
 190                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 191
 192             /* Calculate displacement vector */
 193             dx11             = _mm_sub_pd(ix1,jx1);
 194             dy11             = _mm_sub_pd(iy1,jy1);
 195             dz11             = _mm_sub_pd(iz1,jz1);
 196             dx12             = _mm_sub_pd(ix1,jx2);
 197             dy12             = _mm_sub_pd(iy1,jy2);
 198             dz12             = _mm_sub_pd(iz1,jz2);
 199             dx13             = _mm_sub_pd(ix1,jx3);
 200             dy13             = _mm_sub_pd(iy1,jy3);
 201             dz13             = _mm_sub_pd(iz1,jz3);
 202             dx21             = _mm_sub_pd(ix2,jx1);
 203             dy21             = _mm_sub_pd(iy2,jy1);
 204             dz21             = _mm_sub_pd(iz2,jz1);
 205             dx22             = _mm_sub_pd(ix2,jx2);
 206             dy22             = _mm_sub_pd(iy2,jy2);
 207             dz22             = _mm_sub_pd(iz2,jz2);
 208             dx23             = _mm_sub_pd(ix2,jx3);
 209             dy23             = _mm_sub_pd(iy2,jy3);
 210             dz23             = _mm_sub_pd(iz2,jz3);
 211             dx31             = _mm_sub_pd(ix3,jx1);
 212             dy31             = _mm_sub_pd(iy3,jy1);
 213             dz31             = _mm_sub_pd(iz3,jz1);
 214             dx32             = _mm_sub_pd(ix3,jx2);
 215             dy32             = _mm_sub_pd(iy3,jy2);
 216             dz32             = _mm_sub_pd(iz3,jz2);
 217             dx33             = _mm_sub_pd(ix3,jx3);
 218             dy33             = _mm_sub_pd(iy3,jy3);
 219             dz33             = _mm_sub_pd(iz3,jz3);
 220
 221             /* Calculate squared distance and things based on it */
 222             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 223             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 224             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 225             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 226             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 227             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 228             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 229             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 230             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 231
 232             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 233             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 234             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 235             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 236             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 237             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 238             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 239             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 240             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 241
 242             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 243             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 244             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 245             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 246             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 247             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 248             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 249             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 250             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 251
 252             fjx1             = _mm_setzero_pd();
 253             fjy1             = _mm_setzero_pd();
 254             fjz1             = _mm_setzero_pd();
 255             fjx2             = _mm_setzero_pd();
 256             fjy2             = _mm_setzero_pd();
 257             fjz2             = _mm_setzero_pd();
 258             fjx3             = _mm_setzero_pd();
 259             fjy3             = _mm_setzero_pd();
 260             fjz3             = _mm_setzero_pd();
 261
 262             /**************************
 263              * CALCULATE INTERACTIONS *
 264              **************************/
 265
 266             if (gmx_mm_any_lt(rsq11,rcutoff2))
 267             {
 268
 269             r11              = _mm_mul_pd(rsq11,rinv11);
 270
 271             /* EWALD ELECTROSTATICS */
 272
 273             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 274             ewrt             = _mm_mul_pd(r11,ewtabscale);
 275             ewitab           = _mm_cvttpd_epi32(ewrt);
 276             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 277             ewitab           = _mm_slli_epi32(ewitab,2);
 278             ewtabF           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 279             ewtabD           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 280             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 281             ewtabV           = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,0) +2);
 282             ewtabFn          = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,1) +2);
 283             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 284             felec            = _mm_add_pd(ewtabF,_mm_mul_pd(eweps,ewtabD));
 285             velec            = _mm_sub_pd(ewtabV,_mm_mul_pd(_mm_mul_pd(ewtabhalfspace,eweps),_mm_add_pd(ewtabF,felec)));
 286             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_sub_pd(rinv11,sh_ewald),velec));
 287             felec            = _mm_mul_pd(_mm_mul_pd(qq11,rinv11),_mm_sub_pd(rinvsq11,felec));
 288
 289             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 290
 291             /* Update potential sum for this i atom from the interaction with this j atom. */
 292             velec            = _mm_and_pd(velec,cutoff_mask);
 293             velecsum         = _mm_add_pd(velecsum,velec);
 294
 295             fscal            = felec;
 296
 297             fscal            = _mm_and_pd(fscal,cutoff_mask);
 298
 299             /* Calculate temporary vectorial force */
 300             tx               = _mm_mul_pd(fscal,dx11);
 301             ty               = _mm_mul_pd(fscal,dy11);
 302             tz               = _mm_mul_pd(fscal,dz11);
 303
 304             /* Update vectorial force */
 305             fix1             = _mm_add_pd(fix1,tx);
 306             fiy1             = _mm_add_pd(fiy1,ty);
 307             fiz1             = _mm_add_pd(fiz1,tz);
 308
 309             fjx1             = _mm_add_pd(fjx1,tx);
 310             fjy1             = _mm_add_pd(fjy1,ty);
 311             fjz1             = _mm_add_pd(fjz1,tz);
 312
 313             }
 314
 315             /**************************
 316              * CALCULATE INTERACTIONS *
 317              **************************/
 318
 319             if (gmx_mm_any_lt(rsq12,rcutoff2))
 320             {
 321
 322             r12              = _mm_mul_pd(rsq12,rinv12);
 323
 324             /* EWALD ELECTROSTATICS */
 325
 326             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 327             ewrt             = _mm_mul_pd(r12,ewtabscale);
 328             ewitab           = _mm_cvttpd_epi32(ewrt);
 329             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 330             ewitab           = _mm_slli_epi32(ewitab,2);
 331             ewtabF           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 332             ewtabD           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 333             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 334             ewtabV           = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,0) +2);
 335             ewtabFn          = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,1) +2);
 336             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 337             felec            = _mm_add_pd(ewtabF,_mm_mul_pd(eweps,ewtabD));
 338             velec            = _mm_sub_pd(ewtabV,_mm_mul_pd(_mm_mul_pd(ewtabhalfspace,eweps),_mm_add_pd(ewtabF,felec)));
 339             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_sub_pd(rinv12,sh_ewald),velec));
 340             felec            = _mm_mul_pd(_mm_mul_pd(qq12,rinv12),_mm_sub_pd(rinvsq12,felec));
 341
 342             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 343
 344             /* Update potential sum for this i atom from the interaction with this j atom. */
 345             velec            = _mm_and_pd(velec,cutoff_mask);
 346             velecsum         = _mm_add_pd(velecsum,velec);
 347
 348             fscal            = felec;
 349
 350             fscal            = _mm_and_pd(fscal,cutoff_mask);
 351
 352             /* Calculate temporary vectorial force */
 353             tx               = _mm_mul_pd(fscal,dx12);
 354             ty               = _mm_mul_pd(fscal,dy12);
 355             tz               = _mm_mul_pd(fscal,dz12);
 356
 357             /* Update vectorial force */
 358             fix1             = _mm_add_pd(fix1,tx);
 359             fiy1             = _mm_add_pd(fiy1,ty);
 360             fiz1             = _mm_add_pd(fiz1,tz);
 361
 362             fjx2             = _mm_add_pd(fjx2,tx);
 363             fjy2             = _mm_add_pd(fjy2,ty);
 364             fjz2             = _mm_add_pd(fjz2,tz);
 365
 366             }
 367
 368             /**************************
 369              * CALCULATE INTERACTIONS *
 370              **************************/
 371
 372             if (gmx_mm_any_lt(rsq13,rcutoff2))
 373             {
 374
 375             r13              = _mm_mul_pd(rsq13,rinv13);
 376
 377             /* EWALD ELECTROSTATICS */
 378
 379             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 380             ewrt             = _mm_mul_pd(r13,ewtabscale);
 381             ewitab           = _mm_cvttpd_epi32(ewrt);
 382             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 383             ewitab           = _mm_slli_epi32(ewitab,2);
 384             ewtabF           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 385             ewtabD           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 386             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 387             ewtabV           = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,0) +2);
 388             ewtabFn          = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,1) +2);
 389             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 390             felec            = _mm_add_pd(ewtabF,_mm_mul_pd(eweps,ewtabD));
 391             velec            = _mm_sub_pd(ewtabV,_mm_mul_pd(_mm_mul_pd(ewtabhalfspace,eweps),_mm_add_pd(ewtabF,felec)));
 392             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_sub_pd(rinv13,sh_ewald),velec));
 393             felec            = _mm_mul_pd(_mm_mul_pd(qq13,rinv13),_mm_sub_pd(rinvsq13,felec));
 394
 395             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 396
 397             /* Update potential sum for this i atom from the interaction with this j atom. */
 398             velec            = _mm_and_pd(velec,cutoff_mask);
 399             velecsum         = _mm_add_pd(velecsum,velec);
 400
 401             fscal            = felec;
 402
 403             fscal            = _mm_and_pd(fscal,cutoff_mask);
 404
 405             /* Calculate temporary vectorial force */
 406             tx               = _mm_mul_pd(fscal,dx13);
 407             ty               = _mm_mul_pd(fscal,dy13);
 408             tz               = _mm_mul_pd(fscal,dz13);
 409
 410             /* Update vectorial force */
 411             fix1             = _mm_add_pd(fix1,tx);
 412             fiy1             = _mm_add_pd(fiy1,ty);
 413             fiz1             = _mm_add_pd(fiz1,tz);
 414
 415             fjx3             = _mm_add_pd(fjx3,tx);
 416             fjy3             = _mm_add_pd(fjy3,ty);
 417             fjz3             = _mm_add_pd(fjz3,tz);
 418
 419             }
 420
 421             /**************************
 422              * CALCULATE INTERACTIONS *
 423              **************************/
 424
 425             if (gmx_mm_any_lt(rsq21,rcutoff2))
 426             {
 427
 428             r21              = _mm_mul_pd(rsq21,rinv21);
 429
 430             /* EWALD ELECTROSTATICS */
 431
 432             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 433             ewrt             = _mm_mul_pd(r21,ewtabscale);
 434             ewitab           = _mm_cvttpd_epi32(ewrt);
 435             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 436             ewitab           = _mm_slli_epi32(ewitab,2);
 437             ewtabF           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 438             ewtabD           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 439             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 440             ewtabV           = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,0) +2);
 441             ewtabFn          = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,1) +2);
 442             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 443             felec            = _mm_add_pd(ewtabF,_mm_mul_pd(eweps,ewtabD));
 444             velec            = _mm_sub_pd(ewtabV,_mm_mul_pd(_mm_mul_pd(ewtabhalfspace,eweps),_mm_add_pd(ewtabF,felec)));
 445             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_sub_pd(rinv21,sh_ewald),velec));
 446             felec            = _mm_mul_pd(_mm_mul_pd(qq21,rinv21),_mm_sub_pd(rinvsq21,felec));
 447
 448             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 449
 450             /* Update potential sum for this i atom from the interaction with this j atom. */
 451             velec            = _mm_and_pd(velec,cutoff_mask);
 452             velecsum         = _mm_add_pd(velecsum,velec);
 453
 454             fscal            = felec;
 455
 456             fscal            = _mm_and_pd(fscal,cutoff_mask);
 457
 458             /* Calculate temporary vectorial force */
 459             tx               = _mm_mul_pd(fscal,dx21);
 460             ty               = _mm_mul_pd(fscal,dy21);
 461             tz               = _mm_mul_pd(fscal,dz21);
 462
 463             /* Update vectorial force */
 464             fix2             = _mm_add_pd(fix2,tx);
 465             fiy2             = _mm_add_pd(fiy2,ty);
 466             fiz2             = _mm_add_pd(fiz2,tz);
 467
 468             fjx1             = _mm_add_pd(fjx1,tx);
 469             fjy1             = _mm_add_pd(fjy1,ty);
 470             fjz1             = _mm_add_pd(fjz1,tz);
 471
 472             }
 473
 474             /**************************
 475              * CALCULATE INTERACTIONS *
 476              **************************/
 477
 478             if (gmx_mm_any_lt(rsq22,rcutoff2))
 479             {
 480
 481             r22              = _mm_mul_pd(rsq22,rinv22);
 482
 483             /* EWALD ELECTROSTATICS */
 484
 485             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 486             ewrt             = _mm_mul_pd(r22,ewtabscale);
 487             ewitab           = _mm_cvttpd_epi32(ewrt);
 488             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 489             ewitab           = _mm_slli_epi32(ewitab,2);
 490             ewtabF           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 491             ewtabD           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 492             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 493             ewtabV           = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,0) +2);
 494             ewtabFn          = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,1) +2);
 495             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 496             felec            = _mm_add_pd(ewtabF,_mm_mul_pd(eweps,ewtabD));
 497             velec            = _mm_sub_pd(ewtabV,_mm_mul_pd(_mm_mul_pd(ewtabhalfspace,eweps),_mm_add_pd(ewtabF,felec)));
 498             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_sub_pd(rinv22,sh_ewald),velec));
 499             felec            = _mm_mul_pd(_mm_mul_pd(qq22,rinv22),_mm_sub_pd(rinvsq22,felec));
 500
 501             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 502
 503             /* Update potential sum for this i atom from the interaction with this j atom. */
 504             velec            = _mm_and_pd(velec,cutoff_mask);
 505             velecsum         = _mm_add_pd(velecsum,velec);
 506
 507             fscal            = felec;
 508
 509             fscal            = _mm_and_pd(fscal,cutoff_mask);
 510
 511             /* Calculate temporary vectorial force */
 512             tx               = _mm_mul_pd(fscal,dx22);
 513             ty               = _mm_mul_pd(fscal,dy22);
 514             tz               = _mm_mul_pd(fscal,dz22);
 515
 516             /* Update vectorial force */
 517             fix2             = _mm_add_pd(fix2,tx);
 518             fiy2             = _mm_add_pd(fiy2,ty);
 519             fiz2             = _mm_add_pd(fiz2,tz);
 520
 521             fjx2             = _mm_add_pd(fjx2,tx);
 522             fjy2             = _mm_add_pd(fjy2,ty);
 523             fjz2             = _mm_add_pd(fjz2,tz);
 524
 525             }
 526
 527             /**************************
 528              * CALCULATE INTERACTIONS *
 529              **************************/
 530
 531             if (gmx_mm_any_lt(rsq23,rcutoff2))
 532             {
 533
 534             r23              = _mm_mul_pd(rsq23,rinv23);
 535
 536             /* EWALD ELECTROSTATICS */
 537
 538             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 539             ewrt             = _mm_mul_pd(r23,ewtabscale);
 540             ewitab           = _mm_cvttpd_epi32(ewrt);
 541             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 542             ewitab           = _mm_slli_epi32(ewitab,2);
 543             ewtabF           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 544             ewtabD           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 545             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 546             ewtabV           = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,0) +2);
 547             ewtabFn          = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,1) +2);
 548             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 549             felec            = _mm_add_pd(ewtabF,_mm_mul_pd(eweps,ewtabD));
 550             velec            = _mm_sub_pd(ewtabV,_mm_mul_pd(_mm_mul_pd(ewtabhalfspace,eweps),_mm_add_pd(ewtabF,felec)));
 551             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_sub_pd(rinv23,sh_ewald),velec));
 552             felec            = _mm_mul_pd(_mm_mul_pd(qq23,rinv23),_mm_sub_pd(rinvsq23,felec));
 553
 554             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 555
 556             /* Update potential sum for this i atom from the interaction with this j atom. */
 557             velec            = _mm_and_pd(velec,cutoff_mask);
 558             velecsum         = _mm_add_pd(velecsum,velec);
 559
 560             fscal            = felec;
 561
 562             fscal            = _mm_and_pd(fscal,cutoff_mask);
 563
 564             /* Calculate temporary vectorial force */
 565             tx               = _mm_mul_pd(fscal,dx23);
 566             ty               = _mm_mul_pd(fscal,dy23);
 567             tz               = _mm_mul_pd(fscal,dz23);
 568
 569             /* Update vectorial force */
 570             fix2             = _mm_add_pd(fix2,tx);
 571             fiy2             = _mm_add_pd(fiy2,ty);
 572             fiz2             = _mm_add_pd(fiz2,tz);
 573
 574             fjx3             = _mm_add_pd(fjx3,tx);
 575             fjy3             = _mm_add_pd(fjy3,ty);
 576             fjz3             = _mm_add_pd(fjz3,tz);
 577
 578             }
 579
 580             /**************************
 581              * CALCULATE INTERACTIONS *
 582              **************************/
 583
 584             if (gmx_mm_any_lt(rsq31,rcutoff2))
 585             {
 586
 587             r31              = _mm_mul_pd(rsq31,rinv31);
 588
 589             /* EWALD ELECTROSTATICS */
 590
 591             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 592             ewrt             = _mm_mul_pd(r31,ewtabscale);
 593             ewitab           = _mm_cvttpd_epi32(ewrt);
 594             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 595             ewitab           = _mm_slli_epi32(ewitab,2);
 596             ewtabF           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 597             ewtabD           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 598             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 599             ewtabV           = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,0) +2);
 600             ewtabFn          = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,1) +2);
 601             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 602             felec            = _mm_add_pd(ewtabF,_mm_mul_pd(eweps,ewtabD));
 603             velec            = _mm_sub_pd(ewtabV,_mm_mul_pd(_mm_mul_pd(ewtabhalfspace,eweps),_mm_add_pd(ewtabF,felec)));
 604             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_sub_pd(rinv31,sh_ewald),velec));
 605             felec            = _mm_mul_pd(_mm_mul_pd(qq31,rinv31),_mm_sub_pd(rinvsq31,felec));
 606
 607             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
 608
 609             /* Update potential sum for this i atom from the interaction with this j atom. */
 610             velec            = _mm_and_pd(velec,cutoff_mask);
 611             velecsum         = _mm_add_pd(velecsum,velec);
 612
 613             fscal            = felec;
 614
 615             fscal            = _mm_and_pd(fscal,cutoff_mask);
 616
 617             /* Calculate temporary vectorial force */
 618             tx               = _mm_mul_pd(fscal,dx31);
 619             ty               = _mm_mul_pd(fscal,dy31);
 620             tz               = _mm_mul_pd(fscal,dz31);
 621
 622             /* Update vectorial force */
 623             fix3             = _mm_add_pd(fix3,tx);
 624             fiy3             = _mm_add_pd(fiy3,ty);
 625             fiz3             = _mm_add_pd(fiz3,tz);
 626
 627             fjx1             = _mm_add_pd(fjx1,tx);
 628             fjy1             = _mm_add_pd(fjy1,ty);
 629             fjz1             = _mm_add_pd(fjz1,tz);
 630
 631             }
 632
 633             /**************************
 634              * CALCULATE INTERACTIONS *
 635              **************************/
 636
 637             if (gmx_mm_any_lt(rsq32,rcutoff2))
 638             {
 639
 640             r32              = _mm_mul_pd(rsq32,rinv32);
 641
 642             /* EWALD ELECTROSTATICS */
 643
 644             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 645             ewrt             = _mm_mul_pd(r32,ewtabscale);
 646             ewitab           = _mm_cvttpd_epi32(ewrt);
 647             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 648             ewitab           = _mm_slli_epi32(ewitab,2);
 649             ewtabF           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 650             ewtabD           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 651             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 652             ewtabV           = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,0) +2);
 653             ewtabFn          = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,1) +2);
 654             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 655             felec            = _mm_add_pd(ewtabF,_mm_mul_pd(eweps,ewtabD));
 656             velec            = _mm_sub_pd(ewtabV,_mm_mul_pd(_mm_mul_pd(ewtabhalfspace,eweps),_mm_add_pd(ewtabF,felec)));
 657             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_sub_pd(rinv32,sh_ewald),velec));
 658             felec            = _mm_mul_pd(_mm_mul_pd(qq32,rinv32),_mm_sub_pd(rinvsq32,felec));
 659
 660             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
 661
 662             /* Update potential sum for this i atom from the interaction with this j atom. */
 663             velec            = _mm_and_pd(velec,cutoff_mask);
 664             velecsum         = _mm_add_pd(velecsum,velec);
 665
 666             fscal            = felec;
 667
 668             fscal            = _mm_and_pd(fscal,cutoff_mask);
 669
 670             /* Calculate temporary vectorial force */
 671             tx               = _mm_mul_pd(fscal,dx32);
 672             ty               = _mm_mul_pd(fscal,dy32);
 673             tz               = _mm_mul_pd(fscal,dz32);
 674
 675             /* Update vectorial force */
 676             fix3             = _mm_add_pd(fix3,tx);
 677             fiy3             = _mm_add_pd(fiy3,ty);
 678             fiz3             = _mm_add_pd(fiz3,tz);
 679
 680             fjx2             = _mm_add_pd(fjx2,tx);
 681             fjy2             = _mm_add_pd(fjy2,ty);
 682             fjz2             = _mm_add_pd(fjz2,tz);
 683
 684             }
 685
 686             /**************************
 687              * CALCULATE INTERACTIONS *
 688              **************************/
 689
 690             if (gmx_mm_any_lt(rsq33,rcutoff2))
 691             {
 692
 693             r33              = _mm_mul_pd(rsq33,rinv33);
 694
 695             /* EWALD ELECTROSTATICS */
 696
 697             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 698             ewrt             = _mm_mul_pd(r33,ewtabscale);
 699             ewitab           = _mm_cvttpd_epi32(ewrt);
 700             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 701             ewitab           = _mm_slli_epi32(ewitab,2);
 702             ewtabF           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 703             ewtabD           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,1) );
 704             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 705             ewtabV           = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,0) +2);
 706             ewtabFn          = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,1) +2);
 707             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 708             felec            = _mm_add_pd(ewtabF,_mm_mul_pd(eweps,ewtabD));
 709             velec            = _mm_sub_pd(ewtabV,_mm_mul_pd(_mm_mul_pd(ewtabhalfspace,eweps),_mm_add_pd(ewtabF,felec)));
 710             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_sub_pd(rinv33,sh_ewald),velec));
 711             felec            = _mm_mul_pd(_mm_mul_pd(qq33,rinv33),_mm_sub_pd(rinvsq33,felec));
 712
 713             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
 714
 715             /* Update potential sum for this i atom from the interaction with this j atom. */
 716             velec            = _mm_and_pd(velec,cutoff_mask);
 717             velecsum         = _mm_add_pd(velecsum,velec);
 718
 719             fscal            = felec;
 720
 721             fscal            = _mm_and_pd(fscal,cutoff_mask);
 722
 723             /* Calculate temporary vectorial force */
 724             tx               = _mm_mul_pd(fscal,dx33);
 725             ty               = _mm_mul_pd(fscal,dy33);
 726             tz               = _mm_mul_pd(fscal,dz33);
 727
 728             /* Update vectorial force */
 729             fix3             = _mm_add_pd(fix3,tx);
 730             fiy3             = _mm_add_pd(fiy3,ty);
 731             fiz3             = _mm_add_pd(fiz3,tz);
 732
 733             fjx3             = _mm_add_pd(fjx3,tx);
 734             fjy3             = _mm_add_pd(fjy3,ty);
 735             fjz3             = _mm_add_pd(fjz3,tz);
 736
 737             }
 738
 739             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 740
 741             /* Inner loop uses 414 flops */
 742         }
 743
 744         if(jidx<j_index_end)
 745         {
 746
 747             jnrA             = jjnr[jidx];
 748             j_coord_offsetA  = DIM*jnrA;
 749
 750             /* load j atom coordinates */
 751             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
 752                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
 753
 754             /* Calculate displacement vector */
 755             dx11             = _mm_sub_pd(ix1,jx1);
 756             dy11             = _mm_sub_pd(iy1,jy1);
 757             dz11             = _mm_sub_pd(iz1,jz1);
 758             dx12             = _mm_sub_pd(ix1,jx2);
 759             dy12             = _mm_sub_pd(iy1,jy2);
 760             dz12             = _mm_sub_pd(iz1,jz2);
 761             dx13             = _mm_sub_pd(ix1,jx3);
 762             dy13             = _mm_sub_pd(iy1,jy3);
 763             dz13             = _mm_sub_pd(iz1,jz3);
 764             dx21             = _mm_sub_pd(ix2,jx1);
 765             dy21             = _mm_sub_pd(iy2,jy1);
 766             dz21             = _mm_sub_pd(iz2,jz1);
 767             dx22             = _mm_sub_pd(ix2,jx2);
 768             dy22             = _mm_sub_pd(iy2,jy2);
 769             dz22             = _mm_sub_pd(iz2,jz2);
 770             dx23             = _mm_sub_pd(ix2,jx3);
 771             dy23             = _mm_sub_pd(iy2,jy3);
 772             dz23             = _mm_sub_pd(iz2,jz3);
 773             dx31             = _mm_sub_pd(ix3,jx1);
 774             dy31             = _mm_sub_pd(iy3,jy1);
 775             dz31             = _mm_sub_pd(iz3,jz1);
 776             dx32             = _mm_sub_pd(ix3,jx2);
 777             dy32             = _mm_sub_pd(iy3,jy2);
 778             dz32             = _mm_sub_pd(iz3,jz2);
 779             dx33             = _mm_sub_pd(ix3,jx3);
 780             dy33             = _mm_sub_pd(iy3,jy3);
 781             dz33             = _mm_sub_pd(iz3,jz3);
 782
 783             /* Calculate squared distance and things based on it */
 784             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 785             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 786             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 787             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 788             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 789             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 790             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 791             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 792             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 793
 794             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 795             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 796             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 797             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 798             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 799             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 800             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 801             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 802             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 803
 804             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 805             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 806             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 807             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 808             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 809             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 810             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 811             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 812             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 813
 814             fjx1             = _mm_setzero_pd();
 815             fjy1             = _mm_setzero_pd();
 816             fjz1             = _mm_setzero_pd();
 817             fjx2             = _mm_setzero_pd();
 818             fjy2             = _mm_setzero_pd();
 819             fjz2             = _mm_setzero_pd();
 820             fjx3             = _mm_setzero_pd();
 821             fjy3             = _mm_setzero_pd();
 822             fjz3             = _mm_setzero_pd();
 823
 824             /**************************
 825              * CALCULATE INTERACTIONS *
 826              **************************/
 827
 828             if (gmx_mm_any_lt(rsq11,rcutoff2))
 829             {
 830
 831             r11              = _mm_mul_pd(rsq11,rinv11);
 832
 833             /* EWALD ELECTROSTATICS */
 834
 835             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 836             ewrt             = _mm_mul_pd(r11,ewtabscale);
 837             ewitab           = _mm_cvttpd_epi32(ewrt);
 838             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 839             ewitab           = _mm_slli_epi32(ewitab,2);
 840             ewtabF           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 841             ewtabD           = _mm_setzero_pd();
 842             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 843             ewtabV           = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,0) +2);
 844             ewtabFn          = _mm_setzero_pd();
 845             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 846             felec            = _mm_add_pd(ewtabF,_mm_mul_pd(eweps,ewtabD));
 847             velec            = _mm_sub_pd(ewtabV,_mm_mul_pd(_mm_mul_pd(ewtabhalfspace,eweps),_mm_add_pd(ewtabF,felec)));
 848             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_sub_pd(rinv11,sh_ewald),velec));
 849             felec            = _mm_mul_pd(_mm_mul_pd(qq11,rinv11),_mm_sub_pd(rinvsq11,felec));
 850
 851             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 852
 853             /* Update potential sum for this i atom from the interaction with this j atom. */
 854             velec            = _mm_and_pd(velec,cutoff_mask);
 855             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 856             velecsum         = _mm_add_pd(velecsum,velec);
 857
 858             fscal            = felec;
 859
 860             fscal            = _mm_and_pd(fscal,cutoff_mask);
 861
 862             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 863
 864             /* Calculate temporary vectorial force */
 865             tx               = _mm_mul_pd(fscal,dx11);
 866             ty               = _mm_mul_pd(fscal,dy11);
 867             tz               = _mm_mul_pd(fscal,dz11);
 868
 869             /* Update vectorial force */
 870             fix1             = _mm_add_pd(fix1,tx);
 871             fiy1             = _mm_add_pd(fiy1,ty);
 872             fiz1             = _mm_add_pd(fiz1,tz);
 873
 874             fjx1             = _mm_add_pd(fjx1,tx);
 875             fjy1             = _mm_add_pd(fjy1,ty);
 876             fjz1             = _mm_add_pd(fjz1,tz);
 877
 878             }
 879
 880             /**************************
 881              * CALCULATE INTERACTIONS *
 882              **************************/
 883
 884             if (gmx_mm_any_lt(rsq12,rcutoff2))
 885             {
 886
 887             r12              = _mm_mul_pd(rsq12,rinv12);
 888
 889             /* EWALD ELECTROSTATICS */
 890
 891             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 892             ewrt             = _mm_mul_pd(r12,ewtabscale);
 893             ewitab           = _mm_cvttpd_epi32(ewrt);
 894             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 895             ewitab           = _mm_slli_epi32(ewitab,2);
 896             ewtabF           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 897             ewtabD           = _mm_setzero_pd();
 898             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 899             ewtabV           = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,0) +2);
 900             ewtabFn          = _mm_setzero_pd();
 901             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 902             felec            = _mm_add_pd(ewtabF,_mm_mul_pd(eweps,ewtabD));
 903             velec            = _mm_sub_pd(ewtabV,_mm_mul_pd(_mm_mul_pd(ewtabhalfspace,eweps),_mm_add_pd(ewtabF,felec)));
 904             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_sub_pd(rinv12,sh_ewald),velec));
 905             felec            = _mm_mul_pd(_mm_mul_pd(qq12,rinv12),_mm_sub_pd(rinvsq12,felec));
 906
 907             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 908
 909             /* Update potential sum for this i atom from the interaction with this j atom. */
 910             velec            = _mm_and_pd(velec,cutoff_mask);
 911             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 912             velecsum         = _mm_add_pd(velecsum,velec);
 913
 914             fscal            = felec;
 915
 916             fscal            = _mm_and_pd(fscal,cutoff_mask);
 917
 918             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 919
 920             /* Calculate temporary vectorial force */
 921             tx               = _mm_mul_pd(fscal,dx12);
 922             ty               = _mm_mul_pd(fscal,dy12);
 923             tz               = _mm_mul_pd(fscal,dz12);
 924
 925             /* Update vectorial force */
 926             fix1             = _mm_add_pd(fix1,tx);
 927             fiy1             = _mm_add_pd(fiy1,ty);
 928             fiz1             = _mm_add_pd(fiz1,tz);
 929
 930             fjx2             = _mm_add_pd(fjx2,tx);
 931             fjy2             = _mm_add_pd(fjy2,ty);
 932             fjz2             = _mm_add_pd(fjz2,tz);
 933
 934             }
 935
 936             /**************************
 937              * CALCULATE INTERACTIONS *
 938              **************************/
 939
 940             if (gmx_mm_any_lt(rsq13,rcutoff2))
 941             {
 942
 943             r13              = _mm_mul_pd(rsq13,rinv13);
 944
 945             /* EWALD ELECTROSTATICS */
 946
 947             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
 948             ewrt             = _mm_mul_pd(r13,ewtabscale);
 949             ewitab           = _mm_cvttpd_epi32(ewrt);
 950             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
 951             ewitab           = _mm_slli_epi32(ewitab,2);
 952             ewtabF           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,0) );
 953             ewtabD           = _mm_setzero_pd();
 954             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
 955             ewtabV           = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,0) +2);
 956             ewtabFn          = _mm_setzero_pd();
 957             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
 958             felec            = _mm_add_pd(ewtabF,_mm_mul_pd(eweps,ewtabD));
 959             velec            = _mm_sub_pd(ewtabV,_mm_mul_pd(_mm_mul_pd(ewtabhalfspace,eweps),_mm_add_pd(ewtabF,felec)));
 960             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_sub_pd(rinv13,sh_ewald),velec));
 961             felec            = _mm_mul_pd(_mm_mul_pd(qq13,rinv13),_mm_sub_pd(rinvsq13,felec));
 962
 963             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 964
 965             /* Update potential sum for this i atom from the interaction with this j atom. */
 966             velec            = _mm_and_pd(velec,cutoff_mask);
 967             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 968             velecsum         = _mm_add_pd(velecsum,velec);
 969
 970             fscal            = felec;
 971
 972             fscal            = _mm_and_pd(fscal,cutoff_mask);
 973
 974             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 975
 976             /* Calculate temporary vectorial force */
 977             tx               = _mm_mul_pd(fscal,dx13);
 978             ty               = _mm_mul_pd(fscal,dy13);
 979             tz               = _mm_mul_pd(fscal,dz13);
 980
 981             /* Update vectorial force */
 982             fix1             = _mm_add_pd(fix1,tx);
 983             fiy1             = _mm_add_pd(fiy1,ty);
 984             fiz1             = _mm_add_pd(fiz1,tz);
 985
 986             fjx3             = _mm_add_pd(fjx3,tx);
 987             fjy3             = _mm_add_pd(fjy3,ty);
 988             fjz3             = _mm_add_pd(fjz3,tz);
 989
 990             }
 991
 992             /**************************
 993              * CALCULATE INTERACTIONS *
 994              **************************/
 995
 996             if (gmx_mm_any_lt(rsq21,rcutoff2))
 997             {
 998
 999             r21              = _mm_mul_pd(rsq21,rinv21);
1000
1001             /* EWALD ELECTROSTATICS */
1002
1003             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1004             ewrt             = _mm_mul_pd(r21,ewtabscale);
1005             ewitab           = _mm_cvttpd_epi32(ewrt);
1006             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
1007             ewitab           = _mm_slli_epi32(ewitab,2);
1008             ewtabF           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1009             ewtabD           = _mm_setzero_pd();
1010             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
1011             ewtabV           = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,0) +2);
1012             ewtabFn          = _mm_setzero_pd();
1013             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
1014             felec            = _mm_add_pd(ewtabF,_mm_mul_pd(eweps,ewtabD));
1015             velec            = _mm_sub_pd(ewtabV,_mm_mul_pd(_mm_mul_pd(ewtabhalfspace,eweps),_mm_add_pd(ewtabF,felec)));
1016             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_sub_pd(rinv21,sh_ewald),velec));
1017             felec            = _mm_mul_pd(_mm_mul_pd(qq21,rinv21),_mm_sub_pd(rinvsq21,felec));
1018
1019             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1020
1021             /* Update potential sum for this i atom from the interaction with this j atom. */
1022             velec            = _mm_and_pd(velec,cutoff_mask);
1023             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1024             velecsum         = _mm_add_pd(velecsum,velec);
1025
1026             fscal            = felec;
1027
1028             fscal            = _mm_and_pd(fscal,cutoff_mask);
1029
1030             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1031
1032             /* Calculate temporary vectorial force */
1033             tx               = _mm_mul_pd(fscal,dx21);
1034             ty               = _mm_mul_pd(fscal,dy21);
1035             tz               = _mm_mul_pd(fscal,dz21);
1036
1037             /* Update vectorial force */
1038             fix2             = _mm_add_pd(fix2,tx);
1039             fiy2             = _mm_add_pd(fiy2,ty);
1040             fiz2             = _mm_add_pd(fiz2,tz);
1041
1042             fjx1             = _mm_add_pd(fjx1,tx);
1043             fjy1             = _mm_add_pd(fjy1,ty);
1044             fjz1             = _mm_add_pd(fjz1,tz);
1045
1046             }
1047
1048             /**************************
1049              * CALCULATE INTERACTIONS *
1050              **************************/
1051
1052             if (gmx_mm_any_lt(rsq22,rcutoff2))
1053             {
1054
1055             r22              = _mm_mul_pd(rsq22,rinv22);
1056
1057             /* EWALD ELECTROSTATICS */
1058
1059             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1060             ewrt             = _mm_mul_pd(r22,ewtabscale);
1061             ewitab           = _mm_cvttpd_epi32(ewrt);
1062             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
1063             ewitab           = _mm_slli_epi32(ewitab,2);
1064             ewtabF           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1065             ewtabD           = _mm_setzero_pd();
1066             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
1067             ewtabV           = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,0) +2);
1068             ewtabFn          = _mm_setzero_pd();
1069             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
1070             felec            = _mm_add_pd(ewtabF,_mm_mul_pd(eweps,ewtabD));
1071             velec            = _mm_sub_pd(ewtabV,_mm_mul_pd(_mm_mul_pd(ewtabhalfspace,eweps),_mm_add_pd(ewtabF,felec)));
1072             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_sub_pd(rinv22,sh_ewald),velec));
1073             felec            = _mm_mul_pd(_mm_mul_pd(qq22,rinv22),_mm_sub_pd(rinvsq22,felec));
1074
1075             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1076
1077             /* Update potential sum for this i atom from the interaction with this j atom. */
1078             velec            = _mm_and_pd(velec,cutoff_mask);
1079             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1080             velecsum         = _mm_add_pd(velecsum,velec);
1081
1082             fscal            = felec;
1083
1084             fscal            = _mm_and_pd(fscal,cutoff_mask);
1085
1086             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1087
1088             /* Calculate temporary vectorial force */
1089             tx               = _mm_mul_pd(fscal,dx22);
1090             ty               = _mm_mul_pd(fscal,dy22);
1091             tz               = _mm_mul_pd(fscal,dz22);
1092
1093             /* Update vectorial force */
1094             fix2             = _mm_add_pd(fix2,tx);
1095             fiy2             = _mm_add_pd(fiy2,ty);
1096             fiz2             = _mm_add_pd(fiz2,tz);
1097
1098             fjx2             = _mm_add_pd(fjx2,tx);
1099             fjy2             = _mm_add_pd(fjy2,ty);
1100             fjz2             = _mm_add_pd(fjz2,tz);
1101
1102             }
1103
1104             /**************************
1105              * CALCULATE INTERACTIONS *
1106              **************************/
1107
1108             if (gmx_mm_any_lt(rsq23,rcutoff2))
1109             {
1110
1111             r23              = _mm_mul_pd(rsq23,rinv23);
1112
1113             /* EWALD ELECTROSTATICS */
1114
1115             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1116             ewrt             = _mm_mul_pd(r23,ewtabscale);
1117             ewitab           = _mm_cvttpd_epi32(ewrt);
1118             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
1119             ewitab           = _mm_slli_epi32(ewitab,2);
1120             ewtabF           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1121             ewtabD           = _mm_setzero_pd();
1122             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
1123             ewtabV           = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,0) +2);
1124             ewtabFn          = _mm_setzero_pd();
1125             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
1126             felec            = _mm_add_pd(ewtabF,_mm_mul_pd(eweps,ewtabD));
1127             velec            = _mm_sub_pd(ewtabV,_mm_mul_pd(_mm_mul_pd(ewtabhalfspace,eweps),_mm_add_pd(ewtabF,felec)));
1128             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_sub_pd(rinv23,sh_ewald),velec));
1129             felec            = _mm_mul_pd(_mm_mul_pd(qq23,rinv23),_mm_sub_pd(rinvsq23,felec));
1130
1131             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1132
1133             /* Update potential sum for this i atom from the interaction with this j atom. */
1134             velec            = _mm_and_pd(velec,cutoff_mask);
1135             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1136             velecsum         = _mm_add_pd(velecsum,velec);
1137
1138             fscal            = felec;
1139
1140             fscal            = _mm_and_pd(fscal,cutoff_mask);
1141
1142             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1143
1144             /* Calculate temporary vectorial force */
1145             tx               = _mm_mul_pd(fscal,dx23);
1146             ty               = _mm_mul_pd(fscal,dy23);
1147             tz               = _mm_mul_pd(fscal,dz23);
1148
1149             /* Update vectorial force */
1150             fix2             = _mm_add_pd(fix2,tx);
1151             fiy2             = _mm_add_pd(fiy2,ty);
1152             fiz2             = _mm_add_pd(fiz2,tz);
1153
1154             fjx3             = _mm_add_pd(fjx3,tx);
1155             fjy3             = _mm_add_pd(fjy3,ty);
1156             fjz3             = _mm_add_pd(fjz3,tz);
1157
1158             }
1159
1160             /**************************
1161              * CALCULATE INTERACTIONS *
1162              **************************/
1163
1164             if (gmx_mm_any_lt(rsq31,rcutoff2))
1165             {
1166
1167             r31              = _mm_mul_pd(rsq31,rinv31);
1168
1169             /* EWALD ELECTROSTATICS */
1170
1171             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1172             ewrt             = _mm_mul_pd(r31,ewtabscale);
1173             ewitab           = _mm_cvttpd_epi32(ewrt);
1174             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
1175             ewitab           = _mm_slli_epi32(ewitab,2);
1176             ewtabF           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1177             ewtabD           = _mm_setzero_pd();
1178             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
1179             ewtabV           = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,0) +2);
1180             ewtabFn          = _mm_setzero_pd();
1181             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
1182             felec            = _mm_add_pd(ewtabF,_mm_mul_pd(eweps,ewtabD));
1183             velec            = _mm_sub_pd(ewtabV,_mm_mul_pd(_mm_mul_pd(ewtabhalfspace,eweps),_mm_add_pd(ewtabF,felec)));
1184             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_sub_pd(rinv31,sh_ewald),velec));
1185             felec            = _mm_mul_pd(_mm_mul_pd(qq31,rinv31),_mm_sub_pd(rinvsq31,felec));
1186
1187             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1188
1189             /* Update potential sum for this i atom from the interaction with this j atom. */
1190             velec            = _mm_and_pd(velec,cutoff_mask);
1191             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1192             velecsum         = _mm_add_pd(velecsum,velec);
1193
1194             fscal            = felec;
1195
1196             fscal            = _mm_and_pd(fscal,cutoff_mask);
1197
1198             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1199
1200             /* Calculate temporary vectorial force */
1201             tx               = _mm_mul_pd(fscal,dx31);
1202             ty               = _mm_mul_pd(fscal,dy31);
1203             tz               = _mm_mul_pd(fscal,dz31);
1204
1205             /* Update vectorial force */
1206             fix3             = _mm_add_pd(fix3,tx);
1207             fiy3             = _mm_add_pd(fiy3,ty);
1208             fiz3             = _mm_add_pd(fiz3,tz);
1209
1210             fjx1             = _mm_add_pd(fjx1,tx);
1211             fjy1             = _mm_add_pd(fjy1,ty);
1212             fjz1             = _mm_add_pd(fjz1,tz);
1213
1214             }
1215
1216             /**************************
1217              * CALCULATE INTERACTIONS *
1218              **************************/
1219
1220             if (gmx_mm_any_lt(rsq32,rcutoff2))
1221             {
1222
1223             r32              = _mm_mul_pd(rsq32,rinv32);
1224
1225             /* EWALD ELECTROSTATICS */
1226
1227             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1228             ewrt             = _mm_mul_pd(r32,ewtabscale);
1229             ewitab           = _mm_cvttpd_epi32(ewrt);
1230             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
1231             ewitab           = _mm_slli_epi32(ewitab,2);
1232             ewtabF           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1233             ewtabD           = _mm_setzero_pd();
1234             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
1235             ewtabV           = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,0) +2);
1236             ewtabFn          = _mm_setzero_pd();
1237             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
1238             felec            = _mm_add_pd(ewtabF,_mm_mul_pd(eweps,ewtabD));
1239             velec            = _mm_sub_pd(ewtabV,_mm_mul_pd(_mm_mul_pd(ewtabhalfspace,eweps),_mm_add_pd(ewtabF,felec)));
1240             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_sub_pd(rinv32,sh_ewald),velec));
1241             felec            = _mm_mul_pd(_mm_mul_pd(qq32,rinv32),_mm_sub_pd(rinvsq32,felec));
1242
1243             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1244
1245             /* Update potential sum for this i atom from the interaction with this j atom. */
1246             velec            = _mm_and_pd(velec,cutoff_mask);
1247             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1248             velecsum         = _mm_add_pd(velecsum,velec);
1249
1250             fscal            = felec;
1251
1252             fscal            = _mm_and_pd(fscal,cutoff_mask);
1253
1254             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1255
1256             /* Calculate temporary vectorial force */
1257             tx               = _mm_mul_pd(fscal,dx32);
1258             ty               = _mm_mul_pd(fscal,dy32);
1259             tz               = _mm_mul_pd(fscal,dz32);
1260
1261             /* Update vectorial force */
1262             fix3             = _mm_add_pd(fix3,tx);
1263             fiy3             = _mm_add_pd(fiy3,ty);
1264             fiz3             = _mm_add_pd(fiz3,tz);
1265
1266             fjx2             = _mm_add_pd(fjx2,tx);
1267             fjy2             = _mm_add_pd(fjy2,ty);
1268             fjz2             = _mm_add_pd(fjz2,tz);
1269
1270             }
1271
1272             /**************************
1273              * CALCULATE INTERACTIONS *
1274              **************************/
1275
1276             if (gmx_mm_any_lt(rsq33,rcutoff2))
1277             {
1278
1279             r33              = _mm_mul_pd(rsq33,rinv33);
1280
1281             /* EWALD ELECTROSTATICS */
1282
1283             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1284             ewrt             = _mm_mul_pd(r33,ewtabscale);
1285             ewitab           = _mm_cvttpd_epi32(ewrt);
1286             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
1287             ewitab           = _mm_slli_epi32(ewitab,2);
1288             ewtabF           = _mm_load_pd( ewtab + gmx_mm_extract_epi32(ewitab,0) );
1289             ewtabD           = _mm_setzero_pd();
1290             GMX_MM_TRANSPOSE2_PD(ewtabF,ewtabD);
1291             ewtabV           = _mm_load_sd( ewtab + gmx_mm_extract_epi32(ewitab,0) +2);
1292             ewtabFn          = _mm_setzero_pd();
1293             GMX_MM_TRANSPOSE2_PD(ewtabV,ewtabFn);
1294             felec            = _mm_add_pd(ewtabF,_mm_mul_pd(eweps,ewtabD));
1295             velec            = _mm_sub_pd(ewtabV,_mm_mul_pd(_mm_mul_pd(ewtabhalfspace,eweps),_mm_add_pd(ewtabF,felec)));
1296             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_sub_pd(rinv33,sh_ewald),velec));
1297             felec            = _mm_mul_pd(_mm_mul_pd(qq33,rinv33),_mm_sub_pd(rinvsq33,felec));
1298
1299             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1300
1301             /* Update potential sum for this i atom from the interaction with this j atom. */
1302             velec            = _mm_and_pd(velec,cutoff_mask);
1303             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1304             velecsum         = _mm_add_pd(velecsum,velec);
1305
1306             fscal            = felec;
1307
1308             fscal            = _mm_and_pd(fscal,cutoff_mask);
1309
1310             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1311
1312             /* Calculate temporary vectorial force */
1313             tx               = _mm_mul_pd(fscal,dx33);
1314             ty               = _mm_mul_pd(fscal,dy33);
1315             tz               = _mm_mul_pd(fscal,dz33);
1316
1317             /* Update vectorial force */
1318             fix3             = _mm_add_pd(fix3,tx);
1319             fiy3             = _mm_add_pd(fiy3,ty);
1320             fiz3             = _mm_add_pd(fiz3,tz);
1321
1322             fjx3             = _mm_add_pd(fjx3,tx);
1323             fjy3             = _mm_add_pd(fjy3,ty);
1324             fjz3             = _mm_add_pd(fjz3,tz);
1325
1326             }
1327
1328             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1329
1330             /* Inner loop uses 414 flops */
1331         }
1332
1333         /* End of innermost loop */
1334
1335         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1336                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
1337
1338         ggid                        = gid[iidx];
1339         /* Update potential energies */
1340         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1341
1342         /* Increment number of inner iterations */
1343         inneriter                  += j_index_end - j_index_start;
1344
1345         /* Outer loop uses 19 flops */
1346     }
1347
1348     /* Increment number of outer iterations */
1349     outeriter        += nri;
1350
1351     /* Update outer/inner flops */
1352
1353     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_VF,outeriter*19 + inneriter*414);
1354 }
1355 /*
1356  * Gromacs nonbonded kernel:   nb_kernel_ElecEwSh_VdwNone_GeomW4W4_F_sse4_1_double
1357  * Electrostatics interaction: Ewald
1358  * VdW interaction:            None
1359  * Geometry:                   Water4-Water4
1360  * Calculate force/pot:        Force
1361  */
1362 void
1363 nb_kernel_ElecEwSh_VdwNone_GeomW4W4_F_sse4_1_double
1364                     (t_nblist * gmx_restrict                nlist,
1365                      rvec * gmx_restrict                    xx,
1366                      rvec * gmx_restrict                    ff,
1367                      t_forcerec * gmx_restrict              fr,
1368                      t_mdatoms * gmx_restrict               mdatoms,
1369                      nb_kernel_data_t * gmx_restrict        kernel_data,
1370                      t_nrnb * gmx_restrict                  nrnb)
1371 {
1372     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1373      * just 0 for non-waters.
1374      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1375      * jnr indices corresponding to data put in the four positions in the SIMD register.
1376      */
1377     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1378     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1379     int              jnrA,jnrB;
1380     int              j_coord_offsetA,j_coord_offsetB;
1381     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1382     real             rcutoff_scalar;
1383     real             *shiftvec,*fshift,*x,*f;
1384     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1385     int              vdwioffset1;
1386     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1387     int              vdwioffset2;
1388     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1389     int              vdwioffset3;
1390     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1391     int              vdwjidx1A,vdwjidx1B;
1392     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1393     int              vdwjidx2A,vdwjidx2B;
1394     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1395     int              vdwjidx3A,vdwjidx3B;
1396     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1397     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1398     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1399     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1400     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1401     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1402     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1403     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1404     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1405     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1406     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1407     real             *charge;
1408     __m128i          ewitab;
1409     __m128d          ewtabscale,eweps,sh_ewald,ewrt,ewtabhalfspace,ewtabF,ewtabFn,ewtabD,ewtabV;
1410     real             *ewtab;
1411     __m128d          dummy_mask,cutoff_mask;
1412     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1413     __m128d          one     = _mm_set1_pd(1.0);
1414     __m128d          two     = _mm_set1_pd(2.0);
1415     x                = xx[0];
1416     f                = ff[0];
1417
1418     nri              = nlist->nri;
1419     iinr             = nlist->iinr;
1420     jindex           = nlist->jindex;
1421     jjnr             = nlist->jjnr;
1422     shiftidx         = nlist->shift;
1423     gid              = nlist->gid;
1424     shiftvec         = fr->shift_vec[0];
1425     fshift           = fr->fshift[0];
1426     facel            = _mm_set1_pd(fr->epsfac);
1427     charge           = mdatoms->chargeA;
1428
1429     sh_ewald         = _mm_set1_pd(fr->ic->sh_ewald);
1430     ewtab            = fr->ic->tabq_coul_F;
1431     ewtabscale       = _mm_set1_pd(fr->ic->tabq_scale);
1432     ewtabhalfspace   = _mm_set1_pd(0.5/fr->ic->tabq_scale);
1433
1434     /* Setup water-specific parameters */
1435     inr              = nlist->iinr[0];
1436     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1437     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1438     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1439
1440     jq1              = _mm_set1_pd(charge[inr+1]);
1441     jq2              = _mm_set1_pd(charge[inr+2]);
1442     jq3              = _mm_set1_pd(charge[inr+3]);
1443     qq11             = _mm_mul_pd(iq1,jq1);
1444     qq12             = _mm_mul_pd(iq1,jq2);
1445     qq13             = _mm_mul_pd(iq1,jq3);
1446     qq21             = _mm_mul_pd(iq2,jq1);
1447     qq22             = _mm_mul_pd(iq2,jq2);
1448     qq23             = _mm_mul_pd(iq2,jq3);
1449     qq31             = _mm_mul_pd(iq3,jq1);
1450     qq32             = _mm_mul_pd(iq3,jq2);
1451     qq33             = _mm_mul_pd(iq3,jq3);
1452
1453     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1454     rcutoff_scalar   = fr->rcoulomb;
1455     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1456     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1457
1458     /* Avoid stupid compiler warnings */
1459     jnrA = jnrB = 0;
1460     j_coord_offsetA = 0;
1461     j_coord_offsetB = 0;
1462
1463     outeriter        = 0;
1464     inneriter        = 0;
1465
1466     /* Start outer loop over neighborlists */
1467     for(iidx=0; iidx<nri; iidx++)
1468     {
1469         /* Load shift vector for this list */
1470         i_shift_offset   = DIM*shiftidx[iidx];
1471
1472         /* Load limits for loop over neighbors */
1473         j_index_start    = jindex[iidx];
1474         j_index_end      = jindex[iidx+1];
1475
1476         /* Get outer coordinate index */
1477         inr              = iinr[iidx];
1478         i_coord_offset   = DIM*inr;
1479
1480         /* Load i particle coords and add shift vector */
1481         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset+DIM,
1482                                                  &ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1483
1484         fix1             = _mm_setzero_pd();
1485         fiy1             = _mm_setzero_pd();
1486         fiz1             = _mm_setzero_pd();
1487         fix2             = _mm_setzero_pd();
1488         fiy2             = _mm_setzero_pd();
1489         fiz2             = _mm_setzero_pd();
1490         fix3             = _mm_setzero_pd();
1491         fiy3             = _mm_setzero_pd();
1492         fiz3             = _mm_setzero_pd();
1493
1494         /* Start inner kernel loop */
1495         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1496         {
1497
1498             /* Get j neighbor index, and coordinate index */
1499             jnrA             = jjnr[jidx];
1500             jnrB             = jjnr[jidx+1];
1501             j_coord_offsetA  = DIM*jnrA;
1502             j_coord_offsetB  = DIM*jnrB;
1503
1504             /* load j atom coordinates */
1505             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA+DIM,x+j_coord_offsetB+DIM,
1506                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1507
1508             /* Calculate displacement vector */
1509             dx11             = _mm_sub_pd(ix1,jx1);
1510             dy11             = _mm_sub_pd(iy1,jy1);
1511             dz11             = _mm_sub_pd(iz1,jz1);
1512             dx12             = _mm_sub_pd(ix1,jx2);
1513             dy12             = _mm_sub_pd(iy1,jy2);
1514             dz12             = _mm_sub_pd(iz1,jz2);
1515             dx13             = _mm_sub_pd(ix1,jx3);
1516             dy13             = _mm_sub_pd(iy1,jy3);
1517             dz13             = _mm_sub_pd(iz1,jz3);
1518             dx21             = _mm_sub_pd(ix2,jx1);
1519             dy21             = _mm_sub_pd(iy2,jy1);
1520             dz21             = _mm_sub_pd(iz2,jz1);
1521             dx22             = _mm_sub_pd(ix2,jx2);
1522             dy22             = _mm_sub_pd(iy2,jy2);
1523             dz22             = _mm_sub_pd(iz2,jz2);
1524             dx23             = _mm_sub_pd(ix2,jx3);
1525             dy23             = _mm_sub_pd(iy2,jy3);
1526             dz23             = _mm_sub_pd(iz2,jz3);
1527             dx31             = _mm_sub_pd(ix3,jx1);
1528             dy31             = _mm_sub_pd(iy3,jy1);
1529             dz31             = _mm_sub_pd(iz3,jz1);
1530             dx32             = _mm_sub_pd(ix3,jx2);
1531             dy32             = _mm_sub_pd(iy3,jy2);
1532             dz32             = _mm_sub_pd(iz3,jz2);
1533             dx33             = _mm_sub_pd(ix3,jx3);
1534             dy33             = _mm_sub_pd(iy3,jy3);
1535             dz33             = _mm_sub_pd(iz3,jz3);
1536
1537             /* Calculate squared distance and things based on it */
1538             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1539             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1540             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1541             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1542             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1543             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1544             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1545             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1546             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1547
1548             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1549             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1550             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1551             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1552             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1553             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1554             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1555             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1556             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1557
1558             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1559             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1560             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1561             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1562             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1563             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1564             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1565             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1566             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1567
1568             fjx1             = _mm_setzero_pd();
1569             fjy1             = _mm_setzero_pd();
1570             fjz1             = _mm_setzero_pd();
1571             fjx2             = _mm_setzero_pd();
1572             fjy2             = _mm_setzero_pd();
1573             fjz2             = _mm_setzero_pd();
1574             fjx3             = _mm_setzero_pd();
1575             fjy3             = _mm_setzero_pd();
1576             fjz3             = _mm_setzero_pd();
1577
1578             /**************************
1579              * CALCULATE INTERACTIONS *
1580              **************************/
1581
1582             if (gmx_mm_any_lt(rsq11,rcutoff2))
1583             {
1584
1585             r11              = _mm_mul_pd(rsq11,rinv11);
1586
1587             /* EWALD ELECTROSTATICS */
1588
1589             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1590             ewrt             = _mm_mul_pd(r11,ewtabscale);
1591             ewitab           = _mm_cvttpd_epi32(ewrt);
1592             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
1593             gmx_mm_load_2pair_swizzle_pd(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1594                                          &ewtabF,&ewtabFn);
1595             felec            = _mm_add_pd(_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF),_mm_mul_pd(eweps,ewtabFn));
1596             felec            = _mm_mul_pd(_mm_mul_pd(qq11,rinv11),_mm_sub_pd(rinvsq11,felec));
1597
1598             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1599
1600             fscal            = felec;
1601
1602             fscal            = _mm_and_pd(fscal,cutoff_mask);
1603
1604             /* Calculate temporary vectorial force */
1605             tx               = _mm_mul_pd(fscal,dx11);
1606             ty               = _mm_mul_pd(fscal,dy11);
1607             tz               = _mm_mul_pd(fscal,dz11);
1608
1609             /* Update vectorial force */
1610             fix1             = _mm_add_pd(fix1,tx);
1611             fiy1             = _mm_add_pd(fiy1,ty);
1612             fiz1             = _mm_add_pd(fiz1,tz);
1613
1614             fjx1             = _mm_add_pd(fjx1,tx);
1615             fjy1             = _mm_add_pd(fjy1,ty);
1616             fjz1             = _mm_add_pd(fjz1,tz);
1617
1618             }
1619
1620             /**************************
1621              * CALCULATE INTERACTIONS *
1622              **************************/
1623
1624             if (gmx_mm_any_lt(rsq12,rcutoff2))
1625             {
1626
1627             r12              = _mm_mul_pd(rsq12,rinv12);
1628
1629             /* EWALD ELECTROSTATICS */
1630
1631             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1632             ewrt             = _mm_mul_pd(r12,ewtabscale);
1633             ewitab           = _mm_cvttpd_epi32(ewrt);
1634             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
1635             gmx_mm_load_2pair_swizzle_pd(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1636                                          &ewtabF,&ewtabFn);
1637             felec            = _mm_add_pd(_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF),_mm_mul_pd(eweps,ewtabFn));
1638             felec            = _mm_mul_pd(_mm_mul_pd(qq12,rinv12),_mm_sub_pd(rinvsq12,felec));
1639
1640             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1641
1642             fscal            = felec;
1643
1644             fscal            = _mm_and_pd(fscal,cutoff_mask);
1645
1646             /* Calculate temporary vectorial force */
1647             tx               = _mm_mul_pd(fscal,dx12);
1648             ty               = _mm_mul_pd(fscal,dy12);
1649             tz               = _mm_mul_pd(fscal,dz12);
1650
1651             /* Update vectorial force */
1652             fix1             = _mm_add_pd(fix1,tx);
1653             fiy1             = _mm_add_pd(fiy1,ty);
1654             fiz1             = _mm_add_pd(fiz1,tz);
1655
1656             fjx2             = _mm_add_pd(fjx2,tx);
1657             fjy2             = _mm_add_pd(fjy2,ty);
1658             fjz2             = _mm_add_pd(fjz2,tz);
1659
1660             }
1661
1662             /**************************
1663              * CALCULATE INTERACTIONS *
1664              **************************/
1665
1666             if (gmx_mm_any_lt(rsq13,rcutoff2))
1667             {
1668
1669             r13              = _mm_mul_pd(rsq13,rinv13);
1670
1671             /* EWALD ELECTROSTATICS */
1672
1673             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1674             ewrt             = _mm_mul_pd(r13,ewtabscale);
1675             ewitab           = _mm_cvttpd_epi32(ewrt);
1676             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
1677             gmx_mm_load_2pair_swizzle_pd(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1678                                          &ewtabF,&ewtabFn);
1679             felec            = _mm_add_pd(_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF),_mm_mul_pd(eweps,ewtabFn));
1680             felec            = _mm_mul_pd(_mm_mul_pd(qq13,rinv13),_mm_sub_pd(rinvsq13,felec));
1681
1682             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1683
1684             fscal            = felec;
1685
1686             fscal            = _mm_and_pd(fscal,cutoff_mask);
1687
1688             /* Calculate temporary vectorial force */
1689             tx               = _mm_mul_pd(fscal,dx13);
1690             ty               = _mm_mul_pd(fscal,dy13);
1691             tz               = _mm_mul_pd(fscal,dz13);
1692
1693             /* Update vectorial force */
1694             fix1             = _mm_add_pd(fix1,tx);
1695             fiy1             = _mm_add_pd(fiy1,ty);
1696             fiz1             = _mm_add_pd(fiz1,tz);
1697
1698             fjx3             = _mm_add_pd(fjx3,tx);
1699             fjy3             = _mm_add_pd(fjy3,ty);
1700             fjz3             = _mm_add_pd(fjz3,tz);
1701
1702             }
1703
1704             /**************************
1705              * CALCULATE INTERACTIONS *
1706              **************************/
1707
1708             if (gmx_mm_any_lt(rsq21,rcutoff2))
1709             {
1710
1711             r21              = _mm_mul_pd(rsq21,rinv21);
1712
1713             /* EWALD ELECTROSTATICS */
1714
1715             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1716             ewrt             = _mm_mul_pd(r21,ewtabscale);
1717             ewitab           = _mm_cvttpd_epi32(ewrt);
1718             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
1719             gmx_mm_load_2pair_swizzle_pd(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1720                                          &ewtabF,&ewtabFn);
1721             felec            = _mm_add_pd(_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF),_mm_mul_pd(eweps,ewtabFn));
1722             felec            = _mm_mul_pd(_mm_mul_pd(qq21,rinv21),_mm_sub_pd(rinvsq21,felec));
1723
1724             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1725
1726             fscal            = felec;
1727
1728             fscal            = _mm_and_pd(fscal,cutoff_mask);
1729
1730             /* Calculate temporary vectorial force */
1731             tx               = _mm_mul_pd(fscal,dx21);
1732             ty               = _mm_mul_pd(fscal,dy21);
1733             tz               = _mm_mul_pd(fscal,dz21);
1734
1735             /* Update vectorial force */
1736             fix2             = _mm_add_pd(fix2,tx);
1737             fiy2             = _mm_add_pd(fiy2,ty);
1738             fiz2             = _mm_add_pd(fiz2,tz);
1739
1740             fjx1             = _mm_add_pd(fjx1,tx);
1741             fjy1             = _mm_add_pd(fjy1,ty);
1742             fjz1             = _mm_add_pd(fjz1,tz);
1743
1744             }
1745
1746             /**************************
1747              * CALCULATE INTERACTIONS *
1748              **************************/
1749
1750             if (gmx_mm_any_lt(rsq22,rcutoff2))
1751             {
1752
1753             r22              = _mm_mul_pd(rsq22,rinv22);
1754
1755             /* EWALD ELECTROSTATICS */
1756
1757             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1758             ewrt             = _mm_mul_pd(r22,ewtabscale);
1759             ewitab           = _mm_cvttpd_epi32(ewrt);
1760             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
1761             gmx_mm_load_2pair_swizzle_pd(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1762                                          &ewtabF,&ewtabFn);
1763             felec            = _mm_add_pd(_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF),_mm_mul_pd(eweps,ewtabFn));
1764             felec            = _mm_mul_pd(_mm_mul_pd(qq22,rinv22),_mm_sub_pd(rinvsq22,felec));
1765
1766             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1767
1768             fscal            = felec;
1769
1770             fscal            = _mm_and_pd(fscal,cutoff_mask);
1771
1772             /* Calculate temporary vectorial force */
1773             tx               = _mm_mul_pd(fscal,dx22);
1774             ty               = _mm_mul_pd(fscal,dy22);
1775             tz               = _mm_mul_pd(fscal,dz22);
1776
1777             /* Update vectorial force */
1778             fix2             = _mm_add_pd(fix2,tx);
1779             fiy2             = _mm_add_pd(fiy2,ty);
1780             fiz2             = _mm_add_pd(fiz2,tz);
1781
1782             fjx2             = _mm_add_pd(fjx2,tx);
1783             fjy2             = _mm_add_pd(fjy2,ty);
1784             fjz2             = _mm_add_pd(fjz2,tz);
1785
1786             }
1787
1788             /**************************
1789              * CALCULATE INTERACTIONS *
1790              **************************/
1791
1792             if (gmx_mm_any_lt(rsq23,rcutoff2))
1793             {
1794
1795             r23              = _mm_mul_pd(rsq23,rinv23);
1796
1797             /* EWALD ELECTROSTATICS */
1798
1799             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1800             ewrt             = _mm_mul_pd(r23,ewtabscale);
1801             ewitab           = _mm_cvttpd_epi32(ewrt);
1802             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
1803             gmx_mm_load_2pair_swizzle_pd(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1804                                          &ewtabF,&ewtabFn);
1805             felec            = _mm_add_pd(_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF),_mm_mul_pd(eweps,ewtabFn));
1806             felec            = _mm_mul_pd(_mm_mul_pd(qq23,rinv23),_mm_sub_pd(rinvsq23,felec));
1807
1808             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1809
1810             fscal            = felec;
1811
1812             fscal            = _mm_and_pd(fscal,cutoff_mask);
1813
1814             /* Calculate temporary vectorial force */
1815             tx               = _mm_mul_pd(fscal,dx23);
1816             ty               = _mm_mul_pd(fscal,dy23);
1817             tz               = _mm_mul_pd(fscal,dz23);
1818
1819             /* Update vectorial force */
1820             fix2             = _mm_add_pd(fix2,tx);
1821             fiy2             = _mm_add_pd(fiy2,ty);
1822             fiz2             = _mm_add_pd(fiz2,tz);
1823
1824             fjx3             = _mm_add_pd(fjx3,tx);
1825             fjy3             = _mm_add_pd(fjy3,ty);
1826             fjz3             = _mm_add_pd(fjz3,tz);
1827
1828             }
1829
1830             /**************************
1831              * CALCULATE INTERACTIONS *
1832              **************************/
1833
1834             if (gmx_mm_any_lt(rsq31,rcutoff2))
1835             {
1836
1837             r31              = _mm_mul_pd(rsq31,rinv31);
1838
1839             /* EWALD ELECTROSTATICS */
1840
1841             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1842             ewrt             = _mm_mul_pd(r31,ewtabscale);
1843             ewitab           = _mm_cvttpd_epi32(ewrt);
1844             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
1845             gmx_mm_load_2pair_swizzle_pd(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1846                                          &ewtabF,&ewtabFn);
1847             felec            = _mm_add_pd(_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF),_mm_mul_pd(eweps,ewtabFn));
1848             felec            = _mm_mul_pd(_mm_mul_pd(qq31,rinv31),_mm_sub_pd(rinvsq31,felec));
1849
1850             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1851
1852             fscal            = felec;
1853
1854             fscal            = _mm_and_pd(fscal,cutoff_mask);
1855
1856             /* Calculate temporary vectorial force */
1857             tx               = _mm_mul_pd(fscal,dx31);
1858             ty               = _mm_mul_pd(fscal,dy31);
1859             tz               = _mm_mul_pd(fscal,dz31);
1860
1861             /* Update vectorial force */
1862             fix3             = _mm_add_pd(fix3,tx);
1863             fiy3             = _mm_add_pd(fiy3,ty);
1864             fiz3             = _mm_add_pd(fiz3,tz);
1865
1866             fjx1             = _mm_add_pd(fjx1,tx);
1867             fjy1             = _mm_add_pd(fjy1,ty);
1868             fjz1             = _mm_add_pd(fjz1,tz);
1869
1870             }
1871
1872             /**************************
1873              * CALCULATE INTERACTIONS *
1874              **************************/
1875
1876             if (gmx_mm_any_lt(rsq32,rcutoff2))
1877             {
1878
1879             r32              = _mm_mul_pd(rsq32,rinv32);
1880
1881             /* EWALD ELECTROSTATICS */
1882
1883             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1884             ewrt             = _mm_mul_pd(r32,ewtabscale);
1885             ewitab           = _mm_cvttpd_epi32(ewrt);
1886             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
1887             gmx_mm_load_2pair_swizzle_pd(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1888                                          &ewtabF,&ewtabFn);
1889             felec            = _mm_add_pd(_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF),_mm_mul_pd(eweps,ewtabFn));
1890             felec            = _mm_mul_pd(_mm_mul_pd(qq32,rinv32),_mm_sub_pd(rinvsq32,felec));
1891
1892             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1893
1894             fscal            = felec;
1895
1896             fscal            = _mm_and_pd(fscal,cutoff_mask);
1897
1898             /* Calculate temporary vectorial force */
1899             tx               = _mm_mul_pd(fscal,dx32);
1900             ty               = _mm_mul_pd(fscal,dy32);
1901             tz               = _mm_mul_pd(fscal,dz32);
1902
1903             /* Update vectorial force */
1904             fix3             = _mm_add_pd(fix3,tx);
1905             fiy3             = _mm_add_pd(fiy3,ty);
1906             fiz3             = _mm_add_pd(fiz3,tz);
1907
1908             fjx2             = _mm_add_pd(fjx2,tx);
1909             fjy2             = _mm_add_pd(fjy2,ty);
1910             fjz2             = _mm_add_pd(fjz2,tz);
1911
1912             }
1913
1914             /**************************
1915              * CALCULATE INTERACTIONS *
1916              **************************/
1917
1918             if (gmx_mm_any_lt(rsq33,rcutoff2))
1919             {
1920
1921             r33              = _mm_mul_pd(rsq33,rinv33);
1922
1923             /* EWALD ELECTROSTATICS */
1924
1925             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
1926             ewrt             = _mm_mul_pd(r33,ewtabscale);
1927             ewitab           = _mm_cvttpd_epi32(ewrt);
1928             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
1929             gmx_mm_load_2pair_swizzle_pd(ewtab+gmx_mm_extract_epi32(ewitab,0),ewtab+gmx_mm_extract_epi32(ewitab,1),
1930                                          &ewtabF,&ewtabFn);
1931             felec            = _mm_add_pd(_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF),_mm_mul_pd(eweps,ewtabFn));
1932             felec            = _mm_mul_pd(_mm_mul_pd(qq33,rinv33),_mm_sub_pd(rinvsq33,felec));
1933
1934             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1935
1936             fscal            = felec;
1937
1938             fscal            = _mm_and_pd(fscal,cutoff_mask);
1939
1940             /* Calculate temporary vectorial force */
1941             tx               = _mm_mul_pd(fscal,dx33);
1942             ty               = _mm_mul_pd(fscal,dy33);
1943             tz               = _mm_mul_pd(fscal,dz33);
1944
1945             /* Update vectorial force */
1946             fix3             = _mm_add_pd(fix3,tx);
1947             fiy3             = _mm_add_pd(fiy3,ty);
1948             fiz3             = _mm_add_pd(fiz3,tz);
1949
1950             fjx3             = _mm_add_pd(fjx3,tx);
1951             fjy3             = _mm_add_pd(fjy3,ty);
1952             fjz3             = _mm_add_pd(fjz3,tz);
1953
1954             }
1955
1956             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA+DIM,f+j_coord_offsetB+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1957
1958             /* Inner loop uses 351 flops */
1959         }
1960
1961         if(jidx<j_index_end)
1962         {
1963
1964             jnrA             = jjnr[jidx];
1965             j_coord_offsetA  = DIM*jnrA;
1966
1967             /* load j atom coordinates */
1968             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA+DIM,
1969                                               &jx1,&jy1,&jz1,&jx2,&jy2,&jz2,&jx3,&jy3,&jz3);
1970
1971             /* Calculate displacement vector */
1972             dx11             = _mm_sub_pd(ix1,jx1);
1973             dy11             = _mm_sub_pd(iy1,jy1);
1974             dz11             = _mm_sub_pd(iz1,jz1);
1975             dx12             = _mm_sub_pd(ix1,jx2);
1976             dy12             = _mm_sub_pd(iy1,jy2);
1977             dz12             = _mm_sub_pd(iz1,jz2);
1978             dx13             = _mm_sub_pd(ix1,jx3);
1979             dy13             = _mm_sub_pd(iy1,jy3);
1980             dz13             = _mm_sub_pd(iz1,jz3);
1981             dx21             = _mm_sub_pd(ix2,jx1);
1982             dy21             = _mm_sub_pd(iy2,jy1);
1983             dz21             = _mm_sub_pd(iz2,jz1);
1984             dx22             = _mm_sub_pd(ix2,jx2);
1985             dy22             = _mm_sub_pd(iy2,jy2);
1986             dz22             = _mm_sub_pd(iz2,jz2);
1987             dx23             = _mm_sub_pd(ix2,jx3);
1988             dy23             = _mm_sub_pd(iy2,jy3);
1989             dz23             = _mm_sub_pd(iz2,jz3);
1990             dx31             = _mm_sub_pd(ix3,jx1);
1991             dy31             = _mm_sub_pd(iy3,jy1);
1992             dz31             = _mm_sub_pd(iz3,jz1);
1993             dx32             = _mm_sub_pd(ix3,jx2);
1994             dy32             = _mm_sub_pd(iy3,jy2);
1995             dz32             = _mm_sub_pd(iz3,jz2);
1996             dx33             = _mm_sub_pd(ix3,jx3);
1997             dy33             = _mm_sub_pd(iy3,jy3);
1998             dz33             = _mm_sub_pd(iz3,jz3);
1999
2000             /* Calculate squared distance and things based on it */
2001             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
2002             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
2003             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
2004             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
2005             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
2006             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
2007             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
2008             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
2009             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
2010
2011             rinv11           = gmx_mm_invsqrt_pd(rsq11);
2012             rinv12           = gmx_mm_invsqrt_pd(rsq12);
2013             rinv13           = gmx_mm_invsqrt_pd(rsq13);
2014             rinv21           = gmx_mm_invsqrt_pd(rsq21);
2015             rinv22           = gmx_mm_invsqrt_pd(rsq22);
2016             rinv23           = gmx_mm_invsqrt_pd(rsq23);
2017             rinv31           = gmx_mm_invsqrt_pd(rsq31);
2018             rinv32           = gmx_mm_invsqrt_pd(rsq32);
2019             rinv33           = gmx_mm_invsqrt_pd(rsq33);
2020
2021             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
2022             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
2023             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
2024             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
2025             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
2026             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
2027             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
2028             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
2029             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
2030
2031             fjx1             = _mm_setzero_pd();
2032             fjy1             = _mm_setzero_pd();
2033             fjz1             = _mm_setzero_pd();
2034             fjx2             = _mm_setzero_pd();
2035             fjy2             = _mm_setzero_pd();
2036             fjz2             = _mm_setzero_pd();
2037             fjx3             = _mm_setzero_pd();
2038             fjy3             = _mm_setzero_pd();
2039             fjz3             = _mm_setzero_pd();
2040
2041             /**************************
2042              * CALCULATE INTERACTIONS *
2043              **************************/
2044
2045             if (gmx_mm_any_lt(rsq11,rcutoff2))
2046             {
2047
2048             r11              = _mm_mul_pd(rsq11,rinv11);
2049
2050             /* EWALD ELECTROSTATICS */
2051
2052             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2053             ewrt             = _mm_mul_pd(r11,ewtabscale);
2054             ewitab           = _mm_cvttpd_epi32(ewrt);
2055             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
2056             gmx_mm_load_1pair_swizzle_pd(ewtab+gmx_mm_extract_epi32(ewitab,0),&ewtabF,&ewtabFn);
2057             felec            = _mm_add_pd(_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF),_mm_mul_pd(eweps,ewtabFn));
2058             felec            = _mm_mul_pd(_mm_mul_pd(qq11,rinv11),_mm_sub_pd(rinvsq11,felec));
2059
2060             cutoff_mask      = _mm_cmplt_pd(rsq11,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,dx11);
2070             ty               = _mm_mul_pd(fscal,dy11);
2071             tz               = _mm_mul_pd(fscal,dz11);
2072
2073             /* Update vectorial force */
2074             fix1             = _mm_add_pd(fix1,tx);
2075             fiy1             = _mm_add_pd(fiy1,ty);
2076             fiz1             = _mm_add_pd(fiz1,tz);
2077
2078             fjx1             = _mm_add_pd(fjx1,tx);
2079             fjy1             = _mm_add_pd(fjy1,ty);
2080             fjz1             = _mm_add_pd(fjz1,tz);
2081
2082             }
2083
2084             /**************************
2085              * CALCULATE INTERACTIONS *
2086              **************************/
2087
2088             if (gmx_mm_any_lt(rsq12,rcutoff2))
2089             {
2090
2091             r12              = _mm_mul_pd(rsq12,rinv12);
2092
2093             /* EWALD ELECTROSTATICS */
2094
2095             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2096             ewrt             = _mm_mul_pd(r12,ewtabscale);
2097             ewitab           = _mm_cvttpd_epi32(ewrt);
2098             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
2099             gmx_mm_load_1pair_swizzle_pd(ewtab+gmx_mm_extract_epi32(ewitab,0),&ewtabF,&ewtabFn);
2100             felec            = _mm_add_pd(_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF),_mm_mul_pd(eweps,ewtabFn));
2101             felec            = _mm_mul_pd(_mm_mul_pd(qq12,rinv12),_mm_sub_pd(rinvsq12,felec));
2102
2103             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
2104
2105             fscal            = felec;
2106
2107             fscal            = _mm_and_pd(fscal,cutoff_mask);
2108
2109             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2110
2111             /* Calculate temporary vectorial force */
2112             tx               = _mm_mul_pd(fscal,dx12);
2113             ty               = _mm_mul_pd(fscal,dy12);
2114             tz               = _mm_mul_pd(fscal,dz12);
2115
2116             /* Update vectorial force */
2117             fix1             = _mm_add_pd(fix1,tx);
2118             fiy1             = _mm_add_pd(fiy1,ty);
2119             fiz1             = _mm_add_pd(fiz1,tz);
2120
2121             fjx2             = _mm_add_pd(fjx2,tx);
2122             fjy2             = _mm_add_pd(fjy2,ty);
2123             fjz2             = _mm_add_pd(fjz2,tz);
2124
2125             }
2126
2127             /**************************
2128              * CALCULATE INTERACTIONS *
2129              **************************/
2130
2131             if (gmx_mm_any_lt(rsq13,rcutoff2))
2132             {
2133
2134             r13              = _mm_mul_pd(rsq13,rinv13);
2135
2136             /* EWALD ELECTROSTATICS */
2137
2138             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2139             ewrt             = _mm_mul_pd(r13,ewtabscale);
2140             ewitab           = _mm_cvttpd_epi32(ewrt);
2141             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
2142             gmx_mm_load_1pair_swizzle_pd(ewtab+gmx_mm_extract_epi32(ewitab,0),&ewtabF,&ewtabFn);
2143             felec            = _mm_add_pd(_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF),_mm_mul_pd(eweps,ewtabFn));
2144             felec            = _mm_mul_pd(_mm_mul_pd(qq13,rinv13),_mm_sub_pd(rinvsq13,felec));
2145
2146             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
2147
2148             fscal            = felec;
2149
2150             fscal            = _mm_and_pd(fscal,cutoff_mask);
2151
2152             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2153
2154             /* Calculate temporary vectorial force */
2155             tx               = _mm_mul_pd(fscal,dx13);
2156             ty               = _mm_mul_pd(fscal,dy13);
2157             tz               = _mm_mul_pd(fscal,dz13);
2158
2159             /* Update vectorial force */
2160             fix1             = _mm_add_pd(fix1,tx);
2161             fiy1             = _mm_add_pd(fiy1,ty);
2162             fiz1             = _mm_add_pd(fiz1,tz);
2163
2164             fjx3             = _mm_add_pd(fjx3,tx);
2165             fjy3             = _mm_add_pd(fjy3,ty);
2166             fjz3             = _mm_add_pd(fjz3,tz);
2167
2168             }
2169
2170             /**************************
2171              * CALCULATE INTERACTIONS *
2172              **************************/
2173
2174             if (gmx_mm_any_lt(rsq21,rcutoff2))
2175             {
2176
2177             r21              = _mm_mul_pd(rsq21,rinv21);
2178
2179             /* EWALD ELECTROSTATICS */
2180
2181             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2182             ewrt             = _mm_mul_pd(r21,ewtabscale);
2183             ewitab           = _mm_cvttpd_epi32(ewrt);
2184             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
2185             gmx_mm_load_1pair_swizzle_pd(ewtab+gmx_mm_extract_epi32(ewitab,0),&ewtabF,&ewtabFn);
2186             felec            = _mm_add_pd(_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF),_mm_mul_pd(eweps,ewtabFn));
2187             felec            = _mm_mul_pd(_mm_mul_pd(qq21,rinv21),_mm_sub_pd(rinvsq21,felec));
2188
2189             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
2190
2191             fscal            = felec;
2192
2193             fscal            = _mm_and_pd(fscal,cutoff_mask);
2194
2195             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2196
2197             /* Calculate temporary vectorial force */
2198             tx               = _mm_mul_pd(fscal,dx21);
2199             ty               = _mm_mul_pd(fscal,dy21);
2200             tz               = _mm_mul_pd(fscal,dz21);
2201
2202             /* Update vectorial force */
2203             fix2             = _mm_add_pd(fix2,tx);
2204             fiy2             = _mm_add_pd(fiy2,ty);
2205             fiz2             = _mm_add_pd(fiz2,tz);
2206
2207             fjx1             = _mm_add_pd(fjx1,tx);
2208             fjy1             = _mm_add_pd(fjy1,ty);
2209             fjz1             = _mm_add_pd(fjz1,tz);
2210
2211             }
2212
2213             /**************************
2214              * CALCULATE INTERACTIONS *
2215              **************************/
2216
2217             if (gmx_mm_any_lt(rsq22,rcutoff2))
2218             {
2219
2220             r22              = _mm_mul_pd(rsq22,rinv22);
2221
2222             /* EWALD ELECTROSTATICS */
2223
2224             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2225             ewrt             = _mm_mul_pd(r22,ewtabscale);
2226             ewitab           = _mm_cvttpd_epi32(ewrt);
2227             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
2228             gmx_mm_load_1pair_swizzle_pd(ewtab+gmx_mm_extract_epi32(ewitab,0),&ewtabF,&ewtabFn);
2229             felec            = _mm_add_pd(_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF),_mm_mul_pd(eweps,ewtabFn));
2230             felec            = _mm_mul_pd(_mm_mul_pd(qq22,rinv22),_mm_sub_pd(rinvsq22,felec));
2231
2232             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
2233
2234             fscal            = felec;
2235
2236             fscal            = _mm_and_pd(fscal,cutoff_mask);
2237
2238             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2239
2240             /* Calculate temporary vectorial force */
2241             tx               = _mm_mul_pd(fscal,dx22);
2242             ty               = _mm_mul_pd(fscal,dy22);
2243             tz               = _mm_mul_pd(fscal,dz22);
2244
2245             /* Update vectorial force */
2246             fix2             = _mm_add_pd(fix2,tx);
2247             fiy2             = _mm_add_pd(fiy2,ty);
2248             fiz2             = _mm_add_pd(fiz2,tz);
2249
2250             fjx2             = _mm_add_pd(fjx2,tx);
2251             fjy2             = _mm_add_pd(fjy2,ty);
2252             fjz2             = _mm_add_pd(fjz2,tz);
2253
2254             }
2255
2256             /**************************
2257              * CALCULATE INTERACTIONS *
2258              **************************/
2259
2260             if (gmx_mm_any_lt(rsq23,rcutoff2))
2261             {
2262
2263             r23              = _mm_mul_pd(rsq23,rinv23);
2264
2265             /* EWALD ELECTROSTATICS */
2266
2267             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2268             ewrt             = _mm_mul_pd(r23,ewtabscale);
2269             ewitab           = _mm_cvttpd_epi32(ewrt);
2270             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
2271             gmx_mm_load_1pair_swizzle_pd(ewtab+gmx_mm_extract_epi32(ewitab,0),&ewtabF,&ewtabFn);
2272             felec            = _mm_add_pd(_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF),_mm_mul_pd(eweps,ewtabFn));
2273             felec            = _mm_mul_pd(_mm_mul_pd(qq23,rinv23),_mm_sub_pd(rinvsq23,felec));
2274
2275             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
2276
2277             fscal            = felec;
2278
2279             fscal            = _mm_and_pd(fscal,cutoff_mask);
2280
2281             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2282
2283             /* Calculate temporary vectorial force */
2284             tx               = _mm_mul_pd(fscal,dx23);
2285             ty               = _mm_mul_pd(fscal,dy23);
2286             tz               = _mm_mul_pd(fscal,dz23);
2287
2288             /* Update vectorial force */
2289             fix2             = _mm_add_pd(fix2,tx);
2290             fiy2             = _mm_add_pd(fiy2,ty);
2291             fiz2             = _mm_add_pd(fiz2,tz);
2292
2293             fjx3             = _mm_add_pd(fjx3,tx);
2294             fjy3             = _mm_add_pd(fjy3,ty);
2295             fjz3             = _mm_add_pd(fjz3,tz);
2296
2297             }
2298
2299             /**************************
2300              * CALCULATE INTERACTIONS *
2301              **************************/
2302
2303             if (gmx_mm_any_lt(rsq31,rcutoff2))
2304             {
2305
2306             r31              = _mm_mul_pd(rsq31,rinv31);
2307
2308             /* EWALD ELECTROSTATICS */
2309
2310             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2311             ewrt             = _mm_mul_pd(r31,ewtabscale);
2312             ewitab           = _mm_cvttpd_epi32(ewrt);
2313             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
2314             gmx_mm_load_1pair_swizzle_pd(ewtab+gmx_mm_extract_epi32(ewitab,0),&ewtabF,&ewtabFn);
2315             felec            = _mm_add_pd(_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF),_mm_mul_pd(eweps,ewtabFn));
2316             felec            = _mm_mul_pd(_mm_mul_pd(qq31,rinv31),_mm_sub_pd(rinvsq31,felec));
2317
2318             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
2319
2320             fscal            = felec;
2321
2322             fscal            = _mm_and_pd(fscal,cutoff_mask);
2323
2324             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2325
2326             /* Calculate temporary vectorial force */
2327             tx               = _mm_mul_pd(fscal,dx31);
2328             ty               = _mm_mul_pd(fscal,dy31);
2329             tz               = _mm_mul_pd(fscal,dz31);
2330
2331             /* Update vectorial force */
2332             fix3             = _mm_add_pd(fix3,tx);
2333             fiy3             = _mm_add_pd(fiy3,ty);
2334             fiz3             = _mm_add_pd(fiz3,tz);
2335
2336             fjx1             = _mm_add_pd(fjx1,tx);
2337             fjy1             = _mm_add_pd(fjy1,ty);
2338             fjz1             = _mm_add_pd(fjz1,tz);
2339
2340             }
2341
2342             /**************************
2343              * CALCULATE INTERACTIONS *
2344              **************************/
2345
2346             if (gmx_mm_any_lt(rsq32,rcutoff2))
2347             {
2348
2349             r32              = _mm_mul_pd(rsq32,rinv32);
2350
2351             /* EWALD ELECTROSTATICS */
2352
2353             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2354             ewrt             = _mm_mul_pd(r32,ewtabscale);
2355             ewitab           = _mm_cvttpd_epi32(ewrt);
2356             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
2357             gmx_mm_load_1pair_swizzle_pd(ewtab+gmx_mm_extract_epi32(ewitab,0),&ewtabF,&ewtabFn);
2358             felec            = _mm_add_pd(_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF),_mm_mul_pd(eweps,ewtabFn));
2359             felec            = _mm_mul_pd(_mm_mul_pd(qq32,rinv32),_mm_sub_pd(rinvsq32,felec));
2360
2361             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
2362
2363             fscal            = felec;
2364
2365             fscal            = _mm_and_pd(fscal,cutoff_mask);
2366
2367             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2368
2369             /* Calculate temporary vectorial force */
2370             tx               = _mm_mul_pd(fscal,dx32);
2371             ty               = _mm_mul_pd(fscal,dy32);
2372             tz               = _mm_mul_pd(fscal,dz32);
2373
2374             /* Update vectorial force */
2375             fix3             = _mm_add_pd(fix3,tx);
2376             fiy3             = _mm_add_pd(fiy3,ty);
2377             fiz3             = _mm_add_pd(fiz3,tz);
2378
2379             fjx2             = _mm_add_pd(fjx2,tx);
2380             fjy2             = _mm_add_pd(fjy2,ty);
2381             fjz2             = _mm_add_pd(fjz2,tz);
2382
2383             }
2384
2385             /**************************
2386              * CALCULATE INTERACTIONS *
2387              **************************/
2388
2389             if (gmx_mm_any_lt(rsq33,rcutoff2))
2390             {
2391
2392             r33              = _mm_mul_pd(rsq33,rinv33);
2393
2394             /* EWALD ELECTROSTATICS */
2395
2396             /* Calculate Ewald table index by multiplying r with scale and truncate to integer */
2397             ewrt             = _mm_mul_pd(r33,ewtabscale);
2398             ewitab           = _mm_cvttpd_epi32(ewrt);
2399             eweps            = _mm_sub_pd(ewrt,_mm_round_pd(ewrt, _MM_FROUND_FLOOR));
2400             gmx_mm_load_1pair_swizzle_pd(ewtab+gmx_mm_extract_epi32(ewitab,0),&ewtabF,&ewtabFn);
2401             felec            = _mm_add_pd(_mm_mul_pd( _mm_sub_pd(one,eweps),ewtabF),_mm_mul_pd(eweps,ewtabFn));
2402             felec            = _mm_mul_pd(_mm_mul_pd(qq33,rinv33),_mm_sub_pd(rinvsq33,felec));
2403
2404             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
2405
2406             fscal            = felec;
2407
2408             fscal            = _mm_and_pd(fscal,cutoff_mask);
2409
2410             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2411
2412             /* Calculate temporary vectorial force */
2413             tx               = _mm_mul_pd(fscal,dx33);
2414             ty               = _mm_mul_pd(fscal,dy33);
2415             tz               = _mm_mul_pd(fscal,dz33);
2416
2417             /* Update vectorial force */
2418             fix3             = _mm_add_pd(fix3,tx);
2419             fiy3             = _mm_add_pd(fiy3,ty);
2420             fiz3             = _mm_add_pd(fiz3,tz);
2421
2422             fjx3             = _mm_add_pd(fjx3,tx);
2423             fjy3             = _mm_add_pd(fjy3,ty);
2424             fjz3             = _mm_add_pd(fjz3,tz);
2425
2426             }
2427
2428             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA+DIM,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2429
2430             /* Inner loop uses 351 flops */
2431         }
2432
2433         /* End of innermost loop */
2434
2435         gmx_mm_update_iforce_3atom_swizzle_pd(fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2436                                               f+i_coord_offset+DIM,fshift+i_shift_offset);
2437
2438         /* Increment number of inner iterations */
2439         inneriter                  += j_index_end - j_index_start;
2440
2441         /* Outer loop uses 18 flops */
2442     }
2443
2444     /* Increment number of outer iterations */
2445     outeriter        += nri;
2446
2447     /* Update outer/inner flops */
2448
2449     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_W4W4_F,outeriter*18 + inneriter*351);
2450 }