src/gmxlib/nonbonded/nb_kernel_sse4_1_double/nb_kernel_ElecCoul_VdwCSTab_GeomW3W3_sse4_1_double.c

   1 /*
   2  * Note: this file was generated by the Gromacs sse4_1_double kernel generator.
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  * Copyright (c) 2001-2012, The GROMACS Development Team
   9  *
  10  * Gromacs is a library for molecular simulation and trajectory analysis,
  11  * written by Erik Lindahl, David van der Spoel, Berk Hess, and others - for
  12  * a full list of developers and information, check out http://www.gromacs.org
  13  *
  14  * This program is free software; you can redistribute it and/or modify it under
  15  * the terms of the GNU Lesser General Public License as published by the Free
  16  * Software Foundation; either version 2 of the License, or (at your option) any
  17  * later version.
  18  *
  19  * To help fund GROMACS development, we humbly ask that you cite
  20  * the papers people have written on it - you can find them on the website.
  21  */
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <math.h>
  27
  28 #include "../nb_kernel.h"
  29 #include "types/simple.h"
  30 #include "vec.h"
  31 #include "nrnb.h"
  32
  33 #include "gmx_math_x86_sse4_1_double.h"
  34 #include "kernelutil_x86_sse4_1_double.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwCSTab_GeomW3W3_VF_sse4_1_double
  38  * Electrostatics interaction: Coulomb
  39  * VdW interaction:            CubicSplineTable
  40  * Geometry:                   Water3-Water3
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecCoul_VdwCSTab_GeomW3W3_VF_sse4_1_double
  45                     (t_nblist * gmx_restrict                nlist,
  46                      rvec * gmx_restrict                    xx,
  47                      rvec * gmx_restrict                    ff,
  48                      t_forcerec * gmx_restrict              fr,
  49                      t_mdatoms * gmx_restrict               mdatoms,
  50                      nb_kernel_data_t * gmx_restrict        kernel_data,
  51                      t_nrnb * gmx_restrict                  nrnb)
  52 {
  53     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  54      * just 0 for non-waters.
  55      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  56      * jnr indices corresponding to data put in the four positions in the SIMD register.
  57      */
  58     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  59     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  60     int              jnrA,jnrB;
  61     int              j_coord_offsetA,j_coord_offsetB;
  62     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  63     real             rcutoff_scalar;
  64     real             *shiftvec,*fshift,*x,*f;
  65     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  66     int              vdwioffset0;
  67     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  68     int              vdwioffset1;
  69     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  70     int              vdwioffset2;
  71     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  72     int              vdwjidx0A,vdwjidx0B;
  73     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  74     int              vdwjidx1A,vdwjidx1B;
  75     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  76     int              vdwjidx2A,vdwjidx2B;
  77     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  78     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  79     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  80     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  81     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  82     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  83     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  84     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  85     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  86     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  87     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  88     real             *charge;
  89     int              nvdwtype;
  90     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  91     int              *vdwtype;
  92     real             *vdwparam;
  93     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
  94     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
  95     __m128i          vfitab;
  96     __m128i          ifour       = _mm_set1_epi32(4);
  97     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
  98     real             *vftab;
  99     __m128d          dummy_mask,cutoff_mask;
 100     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 101     __m128d          one     = _mm_set1_pd(1.0);
 102     __m128d          two     = _mm_set1_pd(2.0);
 103     x                = xx[0];
 104     f                = ff[0];
 105
 106     nri              = nlist->nri;
 107     iinr             = nlist->iinr;
 108     jindex           = nlist->jindex;
 109     jjnr             = nlist->jjnr;
 110     shiftidx         = nlist->shift;
 111     gid              = nlist->gid;
 112     shiftvec         = fr->shift_vec[0];
 113     fshift           = fr->fshift[0];
 114     facel            = _mm_set1_pd(fr->epsfac);
 115     charge           = mdatoms->chargeA;
 116     nvdwtype         = fr->ntype;
 117     vdwparam         = fr->nbfp;
 118     vdwtype          = mdatoms->typeA;
 119
 120     vftab            = kernel_data->table_vdw->data;
 121     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
 122
 123     /* Setup water-specific parameters */
 124     inr              = nlist->iinr[0];
 125     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 126     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 127     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 128     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 129
 130     jq0              = _mm_set1_pd(charge[inr+0]);
 131     jq1              = _mm_set1_pd(charge[inr+1]);
 132     jq2              = _mm_set1_pd(charge[inr+2]);
 133     vdwjidx0A        = 2*vdwtype[inr+0];
 134     qq00             = _mm_mul_pd(iq0,jq0);
 135     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 136     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 137     qq01             = _mm_mul_pd(iq0,jq1);
 138     qq02             = _mm_mul_pd(iq0,jq2);
 139     qq10             = _mm_mul_pd(iq1,jq0);
 140     qq11             = _mm_mul_pd(iq1,jq1);
 141     qq12             = _mm_mul_pd(iq1,jq2);
 142     qq20             = _mm_mul_pd(iq2,jq0);
 143     qq21             = _mm_mul_pd(iq2,jq1);
 144     qq22             = _mm_mul_pd(iq2,jq2);
 145
 146     /* Avoid stupid compiler warnings */
 147     jnrA = jnrB = 0;
 148     j_coord_offsetA = 0;
 149     j_coord_offsetB = 0;
 150
 151     outeriter        = 0;
 152     inneriter        = 0;
 153
 154     /* Start outer loop over neighborlists */
 155     for(iidx=0; iidx<nri; iidx++)
 156     {
 157         /* Load shift vector for this list */
 158         i_shift_offset   = DIM*shiftidx[iidx];
 159
 160         /* Load limits for loop over neighbors */
 161         j_index_start    = jindex[iidx];
 162         j_index_end      = jindex[iidx+1];
 163
 164         /* Get outer coordinate index */
 165         inr              = iinr[iidx];
 166         i_coord_offset   = DIM*inr;
 167
 168         /* Load i particle coords and add shift vector */
 169         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 170                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 171
 172         fix0             = _mm_setzero_pd();
 173         fiy0             = _mm_setzero_pd();
 174         fiz0             = _mm_setzero_pd();
 175         fix1             = _mm_setzero_pd();
 176         fiy1             = _mm_setzero_pd();
 177         fiz1             = _mm_setzero_pd();
 178         fix2             = _mm_setzero_pd();
 179         fiy2             = _mm_setzero_pd();
 180         fiz2             = _mm_setzero_pd();
 181
 182         /* Reset potential sums */
 183         velecsum         = _mm_setzero_pd();
 184         vvdwsum          = _mm_setzero_pd();
 185
 186         /* Start inner kernel loop */
 187         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 188         {
 189
 190             /* Get j neighbor index, and coordinate index */
 191             jnrA             = jjnr[jidx];
 192             jnrB             = jjnr[jidx+1];
 193             j_coord_offsetA  = DIM*jnrA;
 194             j_coord_offsetB  = DIM*jnrB;
 195
 196             /* load j atom coordinates */
 197             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 198                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 199
 200             /* Calculate displacement vector */
 201             dx00             = _mm_sub_pd(ix0,jx0);
 202             dy00             = _mm_sub_pd(iy0,jy0);
 203             dz00             = _mm_sub_pd(iz0,jz0);
 204             dx01             = _mm_sub_pd(ix0,jx1);
 205             dy01             = _mm_sub_pd(iy0,jy1);
 206             dz01             = _mm_sub_pd(iz0,jz1);
 207             dx02             = _mm_sub_pd(ix0,jx2);
 208             dy02             = _mm_sub_pd(iy0,jy2);
 209             dz02             = _mm_sub_pd(iz0,jz2);
 210             dx10             = _mm_sub_pd(ix1,jx0);
 211             dy10             = _mm_sub_pd(iy1,jy0);
 212             dz10             = _mm_sub_pd(iz1,jz0);
 213             dx11             = _mm_sub_pd(ix1,jx1);
 214             dy11             = _mm_sub_pd(iy1,jy1);
 215             dz11             = _mm_sub_pd(iz1,jz1);
 216             dx12             = _mm_sub_pd(ix1,jx2);
 217             dy12             = _mm_sub_pd(iy1,jy2);
 218             dz12             = _mm_sub_pd(iz1,jz2);
 219             dx20             = _mm_sub_pd(ix2,jx0);
 220             dy20             = _mm_sub_pd(iy2,jy0);
 221             dz20             = _mm_sub_pd(iz2,jz0);
 222             dx21             = _mm_sub_pd(ix2,jx1);
 223             dy21             = _mm_sub_pd(iy2,jy1);
 224             dz21             = _mm_sub_pd(iz2,jz1);
 225             dx22             = _mm_sub_pd(ix2,jx2);
 226             dy22             = _mm_sub_pd(iy2,jy2);
 227             dz22             = _mm_sub_pd(iz2,jz2);
 228
 229             /* Calculate squared distance and things based on it */
 230             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 231             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 232             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 233             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 234             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 235             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 236             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 237             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 238             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 239
 240             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 241             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 242             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 243             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 244             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 245             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 246             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 247             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 248             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 249
 250             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 251             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 252             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 253             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 254             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 255             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 256             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 257             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 258             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 259
 260             fjx0             = _mm_setzero_pd();
 261             fjy0             = _mm_setzero_pd();
 262             fjz0             = _mm_setzero_pd();
 263             fjx1             = _mm_setzero_pd();
 264             fjy1             = _mm_setzero_pd();
 265             fjz1             = _mm_setzero_pd();
 266             fjx2             = _mm_setzero_pd();
 267             fjy2             = _mm_setzero_pd();
 268             fjz2             = _mm_setzero_pd();
 269
 270             /**************************
 271              * CALCULATE INTERACTIONS *
 272              **************************/
 273
 274             r00              = _mm_mul_pd(rsq00,rinv00);
 275
 276             /* Calculate table index by multiplying r with table scale and truncate to integer */
 277             rt               = _mm_mul_pd(r00,vftabscale);
 278             vfitab           = _mm_cvttpd_epi32(rt);
 279             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 280             vfitab           = _mm_slli_epi32(vfitab,3);
 281
 282             /* COULOMB ELECTROSTATICS */
 283             velec            = _mm_mul_pd(qq00,rinv00);
 284             felec            = _mm_mul_pd(velec,rinvsq00);
 285
 286             /* CUBIC SPLINE TABLE DISPERSION */
 287             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 288             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
 289             GMX_MM_TRANSPOSE2_PD(Y,F);
 290             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 291             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
 292             GMX_MM_TRANSPOSE2_PD(G,H);
 293             Heps             = _mm_mul_pd(vfeps,H);
 294             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 295             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 296             vvdw6            = _mm_mul_pd(c6_00,VV);
 297             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 298             fvdw6            = _mm_mul_pd(c6_00,FF);
 299
 300             /* CUBIC SPLINE TABLE REPULSION */
 301             vfitab           = _mm_add_epi32(vfitab,ifour);
 302             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 303             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
 304             GMX_MM_TRANSPOSE2_PD(Y,F);
 305             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 306             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
 307             GMX_MM_TRANSPOSE2_PD(G,H);
 308             Heps             = _mm_mul_pd(vfeps,H);
 309             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 310             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 311             vvdw12           = _mm_mul_pd(c12_00,VV);
 312             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 313             fvdw12           = _mm_mul_pd(c12_00,FF);
 314             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 315             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 316
 317             /* Update potential sum for this i atom from the interaction with this j atom. */
 318             velecsum         = _mm_add_pd(velecsum,velec);
 319             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 320
 321             fscal            = _mm_add_pd(felec,fvdw);
 322
 323             /* Calculate temporary vectorial force */
 324             tx               = _mm_mul_pd(fscal,dx00);
 325             ty               = _mm_mul_pd(fscal,dy00);
 326             tz               = _mm_mul_pd(fscal,dz00);
 327
 328             /* Update vectorial force */
 329             fix0             = _mm_add_pd(fix0,tx);
 330             fiy0             = _mm_add_pd(fiy0,ty);
 331             fiz0             = _mm_add_pd(fiz0,tz);
 332
 333             fjx0             = _mm_add_pd(fjx0,tx);
 334             fjy0             = _mm_add_pd(fjy0,ty);
 335             fjz0             = _mm_add_pd(fjz0,tz);
 336
 337             /**************************
 338              * CALCULATE INTERACTIONS *
 339              **************************/
 340
 341             /* COULOMB ELECTROSTATICS */
 342             velec            = _mm_mul_pd(qq01,rinv01);
 343             felec            = _mm_mul_pd(velec,rinvsq01);
 344
 345             /* Update potential sum for this i atom from the interaction with this j atom. */
 346             velecsum         = _mm_add_pd(velecsum,velec);
 347
 348             fscal            = felec;
 349
 350             /* Calculate temporary vectorial force */
 351             tx               = _mm_mul_pd(fscal,dx01);
 352             ty               = _mm_mul_pd(fscal,dy01);
 353             tz               = _mm_mul_pd(fscal,dz01);
 354
 355             /* Update vectorial force */
 356             fix0             = _mm_add_pd(fix0,tx);
 357             fiy0             = _mm_add_pd(fiy0,ty);
 358             fiz0             = _mm_add_pd(fiz0,tz);
 359
 360             fjx1             = _mm_add_pd(fjx1,tx);
 361             fjy1             = _mm_add_pd(fjy1,ty);
 362             fjz1             = _mm_add_pd(fjz1,tz);
 363
 364             /**************************
 365              * CALCULATE INTERACTIONS *
 366              **************************/
 367
 368             /* COULOMB ELECTROSTATICS */
 369             velec            = _mm_mul_pd(qq02,rinv02);
 370             felec            = _mm_mul_pd(velec,rinvsq02);
 371
 372             /* Update potential sum for this i atom from the interaction with this j atom. */
 373             velecsum         = _mm_add_pd(velecsum,velec);
 374
 375             fscal            = felec;
 376
 377             /* Calculate temporary vectorial force */
 378             tx               = _mm_mul_pd(fscal,dx02);
 379             ty               = _mm_mul_pd(fscal,dy02);
 380             tz               = _mm_mul_pd(fscal,dz02);
 381
 382             /* Update vectorial force */
 383             fix0             = _mm_add_pd(fix0,tx);
 384             fiy0             = _mm_add_pd(fiy0,ty);
 385             fiz0             = _mm_add_pd(fiz0,tz);
 386
 387             fjx2             = _mm_add_pd(fjx2,tx);
 388             fjy2             = _mm_add_pd(fjy2,ty);
 389             fjz2             = _mm_add_pd(fjz2,tz);
 390
 391             /**************************
 392              * CALCULATE INTERACTIONS *
 393              **************************/
 394
 395             /* COULOMB ELECTROSTATICS */
 396             velec            = _mm_mul_pd(qq10,rinv10);
 397             felec            = _mm_mul_pd(velec,rinvsq10);
 398
 399             /* Update potential sum for this i atom from the interaction with this j atom. */
 400             velecsum         = _mm_add_pd(velecsum,velec);
 401
 402             fscal            = felec;
 403
 404             /* Calculate temporary vectorial force */
 405             tx               = _mm_mul_pd(fscal,dx10);
 406             ty               = _mm_mul_pd(fscal,dy10);
 407             tz               = _mm_mul_pd(fscal,dz10);
 408
 409             /* Update vectorial force */
 410             fix1             = _mm_add_pd(fix1,tx);
 411             fiy1             = _mm_add_pd(fiy1,ty);
 412             fiz1             = _mm_add_pd(fiz1,tz);
 413
 414             fjx0             = _mm_add_pd(fjx0,tx);
 415             fjy0             = _mm_add_pd(fjy0,ty);
 416             fjz0             = _mm_add_pd(fjz0,tz);
 417
 418             /**************************
 419              * CALCULATE INTERACTIONS *
 420              **************************/
 421
 422             /* COULOMB ELECTROSTATICS */
 423             velec            = _mm_mul_pd(qq11,rinv11);
 424             felec            = _mm_mul_pd(velec,rinvsq11);
 425
 426             /* Update potential sum for this i atom from the interaction with this j atom. */
 427             velecsum         = _mm_add_pd(velecsum,velec);
 428
 429             fscal            = felec;
 430
 431             /* Calculate temporary vectorial force */
 432             tx               = _mm_mul_pd(fscal,dx11);
 433             ty               = _mm_mul_pd(fscal,dy11);
 434             tz               = _mm_mul_pd(fscal,dz11);
 435
 436             /* Update vectorial force */
 437             fix1             = _mm_add_pd(fix1,tx);
 438             fiy1             = _mm_add_pd(fiy1,ty);
 439             fiz1             = _mm_add_pd(fiz1,tz);
 440
 441             fjx1             = _mm_add_pd(fjx1,tx);
 442             fjy1             = _mm_add_pd(fjy1,ty);
 443             fjz1             = _mm_add_pd(fjz1,tz);
 444
 445             /**************************
 446              * CALCULATE INTERACTIONS *
 447              **************************/
 448
 449             /* COULOMB ELECTROSTATICS */
 450             velec            = _mm_mul_pd(qq12,rinv12);
 451             felec            = _mm_mul_pd(velec,rinvsq12);
 452
 453             /* Update potential sum for this i atom from the interaction with this j atom. */
 454             velecsum         = _mm_add_pd(velecsum,velec);
 455
 456             fscal            = felec;
 457
 458             /* Calculate temporary vectorial force */
 459             tx               = _mm_mul_pd(fscal,dx12);
 460             ty               = _mm_mul_pd(fscal,dy12);
 461             tz               = _mm_mul_pd(fscal,dz12);
 462
 463             /* Update vectorial force */
 464             fix1             = _mm_add_pd(fix1,tx);
 465             fiy1             = _mm_add_pd(fiy1,ty);
 466             fiz1             = _mm_add_pd(fiz1,tz);
 467
 468             fjx2             = _mm_add_pd(fjx2,tx);
 469             fjy2             = _mm_add_pd(fjy2,ty);
 470             fjz2             = _mm_add_pd(fjz2,tz);
 471
 472             /**************************
 473              * CALCULATE INTERACTIONS *
 474              **************************/
 475
 476             /* COULOMB ELECTROSTATICS */
 477             velec            = _mm_mul_pd(qq20,rinv20);
 478             felec            = _mm_mul_pd(velec,rinvsq20);
 479
 480             /* Update potential sum for this i atom from the interaction with this j atom. */
 481             velecsum         = _mm_add_pd(velecsum,velec);
 482
 483             fscal            = felec;
 484
 485             /* Calculate temporary vectorial force */
 486             tx               = _mm_mul_pd(fscal,dx20);
 487             ty               = _mm_mul_pd(fscal,dy20);
 488             tz               = _mm_mul_pd(fscal,dz20);
 489
 490             /* Update vectorial force */
 491             fix2             = _mm_add_pd(fix2,tx);
 492             fiy2             = _mm_add_pd(fiy2,ty);
 493             fiz2             = _mm_add_pd(fiz2,tz);
 494
 495             fjx0             = _mm_add_pd(fjx0,tx);
 496             fjy0             = _mm_add_pd(fjy0,ty);
 497             fjz0             = _mm_add_pd(fjz0,tz);
 498
 499             /**************************
 500              * CALCULATE INTERACTIONS *
 501              **************************/
 502
 503             /* COULOMB ELECTROSTATICS */
 504             velec            = _mm_mul_pd(qq21,rinv21);
 505             felec            = _mm_mul_pd(velec,rinvsq21);
 506
 507             /* Update potential sum for this i atom from the interaction with this j atom. */
 508             velecsum         = _mm_add_pd(velecsum,velec);
 509
 510             fscal            = felec;
 511
 512             /* Calculate temporary vectorial force */
 513             tx               = _mm_mul_pd(fscal,dx21);
 514             ty               = _mm_mul_pd(fscal,dy21);
 515             tz               = _mm_mul_pd(fscal,dz21);
 516
 517             /* Update vectorial force */
 518             fix2             = _mm_add_pd(fix2,tx);
 519             fiy2             = _mm_add_pd(fiy2,ty);
 520             fiz2             = _mm_add_pd(fiz2,tz);
 521
 522             fjx1             = _mm_add_pd(fjx1,tx);
 523             fjy1             = _mm_add_pd(fjy1,ty);
 524             fjz1             = _mm_add_pd(fjz1,tz);
 525
 526             /**************************
 527              * CALCULATE INTERACTIONS *
 528              **************************/
 529
 530             /* COULOMB ELECTROSTATICS */
 531             velec            = _mm_mul_pd(qq22,rinv22);
 532             felec            = _mm_mul_pd(velec,rinvsq22);
 533
 534             /* Update potential sum for this i atom from the interaction with this j atom. */
 535             velecsum         = _mm_add_pd(velecsum,velec);
 536
 537             fscal            = felec;
 538
 539             /* Calculate temporary vectorial force */
 540             tx               = _mm_mul_pd(fscal,dx22);
 541             ty               = _mm_mul_pd(fscal,dy22);
 542             tz               = _mm_mul_pd(fscal,dz22);
 543
 544             /* Update vectorial force */
 545             fix2             = _mm_add_pd(fix2,tx);
 546             fiy2             = _mm_add_pd(fiy2,ty);
 547             fiz2             = _mm_add_pd(fiz2,tz);
 548
 549             fjx2             = _mm_add_pd(fjx2,tx);
 550             fjy2             = _mm_add_pd(fjy2,ty);
 551             fjz2             = _mm_add_pd(fjz2,tz);
 552
 553             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 554
 555             /* Inner loop uses 287 flops */
 556         }
 557
 558         if(jidx<j_index_end)
 559         {
 560
 561             jnrA             = jjnr[jidx];
 562             j_coord_offsetA  = DIM*jnrA;
 563
 564             /* load j atom coordinates */
 565             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 566                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 567
 568             /* Calculate displacement vector */
 569             dx00             = _mm_sub_pd(ix0,jx0);
 570             dy00             = _mm_sub_pd(iy0,jy0);
 571             dz00             = _mm_sub_pd(iz0,jz0);
 572             dx01             = _mm_sub_pd(ix0,jx1);
 573             dy01             = _mm_sub_pd(iy0,jy1);
 574             dz01             = _mm_sub_pd(iz0,jz1);
 575             dx02             = _mm_sub_pd(ix0,jx2);
 576             dy02             = _mm_sub_pd(iy0,jy2);
 577             dz02             = _mm_sub_pd(iz0,jz2);
 578             dx10             = _mm_sub_pd(ix1,jx0);
 579             dy10             = _mm_sub_pd(iy1,jy0);
 580             dz10             = _mm_sub_pd(iz1,jz0);
 581             dx11             = _mm_sub_pd(ix1,jx1);
 582             dy11             = _mm_sub_pd(iy1,jy1);
 583             dz11             = _mm_sub_pd(iz1,jz1);
 584             dx12             = _mm_sub_pd(ix1,jx2);
 585             dy12             = _mm_sub_pd(iy1,jy2);
 586             dz12             = _mm_sub_pd(iz1,jz2);
 587             dx20             = _mm_sub_pd(ix2,jx0);
 588             dy20             = _mm_sub_pd(iy2,jy0);
 589             dz20             = _mm_sub_pd(iz2,jz0);
 590             dx21             = _mm_sub_pd(ix2,jx1);
 591             dy21             = _mm_sub_pd(iy2,jy1);
 592             dz21             = _mm_sub_pd(iz2,jz1);
 593             dx22             = _mm_sub_pd(ix2,jx2);
 594             dy22             = _mm_sub_pd(iy2,jy2);
 595             dz22             = _mm_sub_pd(iz2,jz2);
 596
 597             /* Calculate squared distance and things based on it */
 598             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 599             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 600             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 601             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 602             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 603             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 604             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 605             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 606             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 607
 608             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 609             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 610             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 611             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 612             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 613             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 614             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 615             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 616             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 617
 618             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 619             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 620             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 621             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 622             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 623             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 624             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 625             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 626             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 627
 628             fjx0             = _mm_setzero_pd();
 629             fjy0             = _mm_setzero_pd();
 630             fjz0             = _mm_setzero_pd();
 631             fjx1             = _mm_setzero_pd();
 632             fjy1             = _mm_setzero_pd();
 633             fjz1             = _mm_setzero_pd();
 634             fjx2             = _mm_setzero_pd();
 635             fjy2             = _mm_setzero_pd();
 636             fjz2             = _mm_setzero_pd();
 637
 638             /**************************
 639              * CALCULATE INTERACTIONS *
 640              **************************/
 641
 642             r00              = _mm_mul_pd(rsq00,rinv00);
 643
 644             /* Calculate table index by multiplying r with table scale and truncate to integer */
 645             rt               = _mm_mul_pd(r00,vftabscale);
 646             vfitab           = _mm_cvttpd_epi32(rt);
 647             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 648             vfitab           = _mm_slli_epi32(vfitab,3);
 649
 650             /* COULOMB ELECTROSTATICS */
 651             velec            = _mm_mul_pd(qq00,rinv00);
 652             felec            = _mm_mul_pd(velec,rinvsq00);
 653
 654             /* CUBIC SPLINE TABLE DISPERSION */
 655             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 656             F                = _mm_setzero_pd();
 657             GMX_MM_TRANSPOSE2_PD(Y,F);
 658             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 659             H                = _mm_setzero_pd();
 660             GMX_MM_TRANSPOSE2_PD(G,H);
 661             Heps             = _mm_mul_pd(vfeps,H);
 662             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 663             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 664             vvdw6            = _mm_mul_pd(c6_00,VV);
 665             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 666             fvdw6            = _mm_mul_pd(c6_00,FF);
 667
 668             /* CUBIC SPLINE TABLE REPULSION */
 669             vfitab           = _mm_add_epi32(vfitab,ifour);
 670             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
 671             F                = _mm_setzero_pd();
 672             GMX_MM_TRANSPOSE2_PD(Y,F);
 673             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
 674             H                = _mm_setzero_pd();
 675             GMX_MM_TRANSPOSE2_PD(G,H);
 676             Heps             = _mm_mul_pd(vfeps,H);
 677             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
 678             VV               = _mm_add_pd(Y,_mm_mul_pd(vfeps,Fp));
 679             vvdw12           = _mm_mul_pd(c12_00,VV);
 680             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
 681             fvdw12           = _mm_mul_pd(c12_00,FF);
 682             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 683             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 684
 685             /* Update potential sum for this i atom from the interaction with this j atom. */
 686             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 687             velecsum         = _mm_add_pd(velecsum,velec);
 688             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 689             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 690
 691             fscal            = _mm_add_pd(felec,fvdw);
 692
 693             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 694
 695             /* Calculate temporary vectorial force */
 696             tx               = _mm_mul_pd(fscal,dx00);
 697             ty               = _mm_mul_pd(fscal,dy00);
 698             tz               = _mm_mul_pd(fscal,dz00);
 699
 700             /* Update vectorial force */
 701             fix0             = _mm_add_pd(fix0,tx);
 702             fiy0             = _mm_add_pd(fiy0,ty);
 703             fiz0             = _mm_add_pd(fiz0,tz);
 704
 705             fjx0             = _mm_add_pd(fjx0,tx);
 706             fjy0             = _mm_add_pd(fjy0,ty);
 707             fjz0             = _mm_add_pd(fjz0,tz);
 708
 709             /**************************
 710              * CALCULATE INTERACTIONS *
 711              **************************/
 712
 713             /* COULOMB ELECTROSTATICS */
 714             velec            = _mm_mul_pd(qq01,rinv01);
 715             felec            = _mm_mul_pd(velec,rinvsq01);
 716
 717             /* Update potential sum for this i atom from the interaction with this j atom. */
 718             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 719             velecsum         = _mm_add_pd(velecsum,velec);
 720
 721             fscal            = felec;
 722
 723             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 724
 725             /* Calculate temporary vectorial force */
 726             tx               = _mm_mul_pd(fscal,dx01);
 727             ty               = _mm_mul_pd(fscal,dy01);
 728             tz               = _mm_mul_pd(fscal,dz01);
 729
 730             /* Update vectorial force */
 731             fix0             = _mm_add_pd(fix0,tx);
 732             fiy0             = _mm_add_pd(fiy0,ty);
 733             fiz0             = _mm_add_pd(fiz0,tz);
 734
 735             fjx1             = _mm_add_pd(fjx1,tx);
 736             fjy1             = _mm_add_pd(fjy1,ty);
 737             fjz1             = _mm_add_pd(fjz1,tz);
 738
 739             /**************************
 740              * CALCULATE INTERACTIONS *
 741              **************************/
 742
 743             /* COULOMB ELECTROSTATICS */
 744             velec            = _mm_mul_pd(qq02,rinv02);
 745             felec            = _mm_mul_pd(velec,rinvsq02);
 746
 747             /* Update potential sum for this i atom from the interaction with this j atom. */
 748             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 749             velecsum         = _mm_add_pd(velecsum,velec);
 750
 751             fscal            = felec;
 752
 753             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 754
 755             /* Calculate temporary vectorial force */
 756             tx               = _mm_mul_pd(fscal,dx02);
 757             ty               = _mm_mul_pd(fscal,dy02);
 758             tz               = _mm_mul_pd(fscal,dz02);
 759
 760             /* Update vectorial force */
 761             fix0             = _mm_add_pd(fix0,tx);
 762             fiy0             = _mm_add_pd(fiy0,ty);
 763             fiz0             = _mm_add_pd(fiz0,tz);
 764
 765             fjx2             = _mm_add_pd(fjx2,tx);
 766             fjy2             = _mm_add_pd(fjy2,ty);
 767             fjz2             = _mm_add_pd(fjz2,tz);
 768
 769             /**************************
 770              * CALCULATE INTERACTIONS *
 771              **************************/
 772
 773             /* COULOMB ELECTROSTATICS */
 774             velec            = _mm_mul_pd(qq10,rinv10);
 775             felec            = _mm_mul_pd(velec,rinvsq10);
 776
 777             /* Update potential sum for this i atom from the interaction with this j atom. */
 778             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 779             velecsum         = _mm_add_pd(velecsum,velec);
 780
 781             fscal            = felec;
 782
 783             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 784
 785             /* Calculate temporary vectorial force */
 786             tx               = _mm_mul_pd(fscal,dx10);
 787             ty               = _mm_mul_pd(fscal,dy10);
 788             tz               = _mm_mul_pd(fscal,dz10);
 789
 790             /* Update vectorial force */
 791             fix1             = _mm_add_pd(fix1,tx);
 792             fiy1             = _mm_add_pd(fiy1,ty);
 793             fiz1             = _mm_add_pd(fiz1,tz);
 794
 795             fjx0             = _mm_add_pd(fjx0,tx);
 796             fjy0             = _mm_add_pd(fjy0,ty);
 797             fjz0             = _mm_add_pd(fjz0,tz);
 798
 799             /**************************
 800              * CALCULATE INTERACTIONS *
 801              **************************/
 802
 803             /* COULOMB ELECTROSTATICS */
 804             velec            = _mm_mul_pd(qq11,rinv11);
 805             felec            = _mm_mul_pd(velec,rinvsq11);
 806
 807             /* Update potential sum for this i atom from the interaction with this j atom. */
 808             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 809             velecsum         = _mm_add_pd(velecsum,velec);
 810
 811             fscal            = felec;
 812
 813             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 814
 815             /* Calculate temporary vectorial force */
 816             tx               = _mm_mul_pd(fscal,dx11);
 817             ty               = _mm_mul_pd(fscal,dy11);
 818             tz               = _mm_mul_pd(fscal,dz11);
 819
 820             /* Update vectorial force */
 821             fix1             = _mm_add_pd(fix1,tx);
 822             fiy1             = _mm_add_pd(fiy1,ty);
 823             fiz1             = _mm_add_pd(fiz1,tz);
 824
 825             fjx1             = _mm_add_pd(fjx1,tx);
 826             fjy1             = _mm_add_pd(fjy1,ty);
 827             fjz1             = _mm_add_pd(fjz1,tz);
 828
 829             /**************************
 830              * CALCULATE INTERACTIONS *
 831              **************************/
 832
 833             /* COULOMB ELECTROSTATICS */
 834             velec            = _mm_mul_pd(qq12,rinv12);
 835             felec            = _mm_mul_pd(velec,rinvsq12);
 836
 837             /* Update potential sum for this i atom from the interaction with this j atom. */
 838             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 839             velecsum         = _mm_add_pd(velecsum,velec);
 840
 841             fscal            = felec;
 842
 843             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 844
 845             /* Calculate temporary vectorial force */
 846             tx               = _mm_mul_pd(fscal,dx12);
 847             ty               = _mm_mul_pd(fscal,dy12);
 848             tz               = _mm_mul_pd(fscal,dz12);
 849
 850             /* Update vectorial force */
 851             fix1             = _mm_add_pd(fix1,tx);
 852             fiy1             = _mm_add_pd(fiy1,ty);
 853             fiz1             = _mm_add_pd(fiz1,tz);
 854
 855             fjx2             = _mm_add_pd(fjx2,tx);
 856             fjy2             = _mm_add_pd(fjy2,ty);
 857             fjz2             = _mm_add_pd(fjz2,tz);
 858
 859             /**************************
 860              * CALCULATE INTERACTIONS *
 861              **************************/
 862
 863             /* COULOMB ELECTROSTATICS */
 864             velec            = _mm_mul_pd(qq20,rinv20);
 865             felec            = _mm_mul_pd(velec,rinvsq20);
 866
 867             /* Update potential sum for this i atom from the interaction with this j atom. */
 868             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 869             velecsum         = _mm_add_pd(velecsum,velec);
 870
 871             fscal            = felec;
 872
 873             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 874
 875             /* Calculate temporary vectorial force */
 876             tx               = _mm_mul_pd(fscal,dx20);
 877             ty               = _mm_mul_pd(fscal,dy20);
 878             tz               = _mm_mul_pd(fscal,dz20);
 879
 880             /* Update vectorial force */
 881             fix2             = _mm_add_pd(fix2,tx);
 882             fiy2             = _mm_add_pd(fiy2,ty);
 883             fiz2             = _mm_add_pd(fiz2,tz);
 884
 885             fjx0             = _mm_add_pd(fjx0,tx);
 886             fjy0             = _mm_add_pd(fjy0,ty);
 887             fjz0             = _mm_add_pd(fjz0,tz);
 888
 889             /**************************
 890              * CALCULATE INTERACTIONS *
 891              **************************/
 892
 893             /* COULOMB ELECTROSTATICS */
 894             velec            = _mm_mul_pd(qq21,rinv21);
 895             felec            = _mm_mul_pd(velec,rinvsq21);
 896
 897             /* Update potential sum for this i atom from the interaction with this j atom. */
 898             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 899             velecsum         = _mm_add_pd(velecsum,velec);
 900
 901             fscal            = felec;
 902
 903             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 904
 905             /* Calculate temporary vectorial force */
 906             tx               = _mm_mul_pd(fscal,dx21);
 907             ty               = _mm_mul_pd(fscal,dy21);
 908             tz               = _mm_mul_pd(fscal,dz21);
 909
 910             /* Update vectorial force */
 911             fix2             = _mm_add_pd(fix2,tx);
 912             fiy2             = _mm_add_pd(fiy2,ty);
 913             fiz2             = _mm_add_pd(fiz2,tz);
 914
 915             fjx1             = _mm_add_pd(fjx1,tx);
 916             fjy1             = _mm_add_pd(fjy1,ty);
 917             fjz1             = _mm_add_pd(fjz1,tz);
 918
 919             /**************************
 920              * CALCULATE INTERACTIONS *
 921              **************************/
 922
 923             /* COULOMB ELECTROSTATICS */
 924             velec            = _mm_mul_pd(qq22,rinv22);
 925             felec            = _mm_mul_pd(velec,rinvsq22);
 926
 927             /* Update potential sum for this i atom from the interaction with this j atom. */
 928             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 929             velecsum         = _mm_add_pd(velecsum,velec);
 930
 931             fscal            = felec;
 932
 933             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 934
 935             /* Calculate temporary vectorial force */
 936             tx               = _mm_mul_pd(fscal,dx22);
 937             ty               = _mm_mul_pd(fscal,dy22);
 938             tz               = _mm_mul_pd(fscal,dz22);
 939
 940             /* Update vectorial force */
 941             fix2             = _mm_add_pd(fix2,tx);
 942             fiy2             = _mm_add_pd(fiy2,ty);
 943             fiz2             = _mm_add_pd(fiz2,tz);
 944
 945             fjx2             = _mm_add_pd(fjx2,tx);
 946             fjy2             = _mm_add_pd(fjy2,ty);
 947             fjz2             = _mm_add_pd(fjz2,tz);
 948
 949             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 950
 951             /* Inner loop uses 287 flops */
 952         }
 953
 954         /* End of innermost loop */
 955
 956         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
 957                                               f+i_coord_offset,fshift+i_shift_offset);
 958
 959         ggid                        = gid[iidx];
 960         /* Update potential energies */
 961         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 962         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 963
 964         /* Increment number of inner iterations */
 965         inneriter                  += j_index_end - j_index_start;
 966
 967         /* Outer loop uses 20 flops */
 968     }
 969
 970     /* Increment number of outer iterations */
 971     outeriter        += nri;
 972
 973     /* Update outer/inner flops */
 974
 975     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*287);
 976 }
 977 /*
 978  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwCSTab_GeomW3W3_F_sse4_1_double
 979  * Electrostatics interaction: Coulomb
 980  * VdW interaction:            CubicSplineTable
 981  * Geometry:                   Water3-Water3
 982  * Calculate force/pot:        Force
 983  */
 984 void
 985 nb_kernel_ElecCoul_VdwCSTab_GeomW3W3_F_sse4_1_double
 986                     (t_nblist * gmx_restrict                nlist,
 987                      rvec * gmx_restrict                    xx,
 988                      rvec * gmx_restrict                    ff,
 989                      t_forcerec * gmx_restrict              fr,
 990                      t_mdatoms * gmx_restrict               mdatoms,
 991                      nb_kernel_data_t * gmx_restrict        kernel_data,
 992                      t_nrnb * gmx_restrict                  nrnb)
 993 {
 994     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 995      * just 0 for non-waters.
 996      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 997      * jnr indices corresponding to data put in the four positions in the SIMD register.
 998      */
 999     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1000     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1001     int              jnrA,jnrB;
1002     int              j_coord_offsetA,j_coord_offsetB;
1003     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1004     real             rcutoff_scalar;
1005     real             *shiftvec,*fshift,*x,*f;
1006     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1007     int              vdwioffset0;
1008     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1009     int              vdwioffset1;
1010     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1011     int              vdwioffset2;
1012     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1013     int              vdwjidx0A,vdwjidx0B;
1014     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1015     int              vdwjidx1A,vdwjidx1B;
1016     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1017     int              vdwjidx2A,vdwjidx2B;
1018     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1019     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1020     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1021     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1022     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1023     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1024     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1025     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1026     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1027     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1028     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1029     real             *charge;
1030     int              nvdwtype;
1031     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1032     int              *vdwtype;
1033     real             *vdwparam;
1034     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1035     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1036     __m128i          vfitab;
1037     __m128i          ifour       = _mm_set1_epi32(4);
1038     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF;
1039     real             *vftab;
1040     __m128d          dummy_mask,cutoff_mask;
1041     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1042     __m128d          one     = _mm_set1_pd(1.0);
1043     __m128d          two     = _mm_set1_pd(2.0);
1044     x                = xx[0];
1045     f                = ff[0];
1046
1047     nri              = nlist->nri;
1048     iinr             = nlist->iinr;
1049     jindex           = nlist->jindex;
1050     jjnr             = nlist->jjnr;
1051     shiftidx         = nlist->shift;
1052     gid              = nlist->gid;
1053     shiftvec         = fr->shift_vec[0];
1054     fshift           = fr->fshift[0];
1055     facel            = _mm_set1_pd(fr->epsfac);
1056     charge           = mdatoms->chargeA;
1057     nvdwtype         = fr->ntype;
1058     vdwparam         = fr->nbfp;
1059     vdwtype          = mdatoms->typeA;
1060
1061     vftab            = kernel_data->table_vdw->data;
1062     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
1063
1064     /* Setup water-specific parameters */
1065     inr              = nlist->iinr[0];
1066     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
1067     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1068     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1069     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1070
1071     jq0              = _mm_set1_pd(charge[inr+0]);
1072     jq1              = _mm_set1_pd(charge[inr+1]);
1073     jq2              = _mm_set1_pd(charge[inr+2]);
1074     vdwjidx0A        = 2*vdwtype[inr+0];
1075     qq00             = _mm_mul_pd(iq0,jq0);
1076     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1077     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1078     qq01             = _mm_mul_pd(iq0,jq1);
1079     qq02             = _mm_mul_pd(iq0,jq2);
1080     qq10             = _mm_mul_pd(iq1,jq0);
1081     qq11             = _mm_mul_pd(iq1,jq1);
1082     qq12             = _mm_mul_pd(iq1,jq2);
1083     qq20             = _mm_mul_pd(iq2,jq0);
1084     qq21             = _mm_mul_pd(iq2,jq1);
1085     qq22             = _mm_mul_pd(iq2,jq2);
1086
1087     /* Avoid stupid compiler warnings */
1088     jnrA = jnrB = 0;
1089     j_coord_offsetA = 0;
1090     j_coord_offsetB = 0;
1091
1092     outeriter        = 0;
1093     inneriter        = 0;
1094
1095     /* Start outer loop over neighborlists */
1096     for(iidx=0; iidx<nri; iidx++)
1097     {
1098         /* Load shift vector for this list */
1099         i_shift_offset   = DIM*shiftidx[iidx];
1100
1101         /* Load limits for loop over neighbors */
1102         j_index_start    = jindex[iidx];
1103         j_index_end      = jindex[iidx+1];
1104
1105         /* Get outer coordinate index */
1106         inr              = iinr[iidx];
1107         i_coord_offset   = DIM*inr;
1108
1109         /* Load i particle coords and add shift vector */
1110         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1111                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1112
1113         fix0             = _mm_setzero_pd();
1114         fiy0             = _mm_setzero_pd();
1115         fiz0             = _mm_setzero_pd();
1116         fix1             = _mm_setzero_pd();
1117         fiy1             = _mm_setzero_pd();
1118         fiz1             = _mm_setzero_pd();
1119         fix2             = _mm_setzero_pd();
1120         fiy2             = _mm_setzero_pd();
1121         fiz2             = _mm_setzero_pd();
1122
1123         /* Start inner kernel loop */
1124         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1125         {
1126
1127             /* Get j neighbor index, and coordinate index */
1128             jnrA             = jjnr[jidx];
1129             jnrB             = jjnr[jidx+1];
1130             j_coord_offsetA  = DIM*jnrA;
1131             j_coord_offsetB  = DIM*jnrB;
1132
1133             /* load j atom coordinates */
1134             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1135                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1136
1137             /* Calculate displacement vector */
1138             dx00             = _mm_sub_pd(ix0,jx0);
1139             dy00             = _mm_sub_pd(iy0,jy0);
1140             dz00             = _mm_sub_pd(iz0,jz0);
1141             dx01             = _mm_sub_pd(ix0,jx1);
1142             dy01             = _mm_sub_pd(iy0,jy1);
1143             dz01             = _mm_sub_pd(iz0,jz1);
1144             dx02             = _mm_sub_pd(ix0,jx2);
1145             dy02             = _mm_sub_pd(iy0,jy2);
1146             dz02             = _mm_sub_pd(iz0,jz2);
1147             dx10             = _mm_sub_pd(ix1,jx0);
1148             dy10             = _mm_sub_pd(iy1,jy0);
1149             dz10             = _mm_sub_pd(iz1,jz0);
1150             dx11             = _mm_sub_pd(ix1,jx1);
1151             dy11             = _mm_sub_pd(iy1,jy1);
1152             dz11             = _mm_sub_pd(iz1,jz1);
1153             dx12             = _mm_sub_pd(ix1,jx2);
1154             dy12             = _mm_sub_pd(iy1,jy2);
1155             dz12             = _mm_sub_pd(iz1,jz2);
1156             dx20             = _mm_sub_pd(ix2,jx0);
1157             dy20             = _mm_sub_pd(iy2,jy0);
1158             dz20             = _mm_sub_pd(iz2,jz0);
1159             dx21             = _mm_sub_pd(ix2,jx1);
1160             dy21             = _mm_sub_pd(iy2,jy1);
1161             dz21             = _mm_sub_pd(iz2,jz1);
1162             dx22             = _mm_sub_pd(ix2,jx2);
1163             dy22             = _mm_sub_pd(iy2,jy2);
1164             dz22             = _mm_sub_pd(iz2,jz2);
1165
1166             /* Calculate squared distance and things based on it */
1167             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1168             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1169             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1170             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1171             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1172             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1173             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1174             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1175             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1176
1177             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1178             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1179             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1180             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1181             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1182             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1183             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1184             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1185             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1186
1187             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1188             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1189             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1190             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1191             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1192             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1193             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1194             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1195             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1196
1197             fjx0             = _mm_setzero_pd();
1198             fjy0             = _mm_setzero_pd();
1199             fjz0             = _mm_setzero_pd();
1200             fjx1             = _mm_setzero_pd();
1201             fjy1             = _mm_setzero_pd();
1202             fjz1             = _mm_setzero_pd();
1203             fjx2             = _mm_setzero_pd();
1204             fjy2             = _mm_setzero_pd();
1205             fjz2             = _mm_setzero_pd();
1206
1207             /**************************
1208              * CALCULATE INTERACTIONS *
1209              **************************/
1210
1211             r00              = _mm_mul_pd(rsq00,rinv00);
1212
1213             /* Calculate table index by multiplying r with table scale and truncate to integer */
1214             rt               = _mm_mul_pd(r00,vftabscale);
1215             vfitab           = _mm_cvttpd_epi32(rt);
1216             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
1217             vfitab           = _mm_slli_epi32(vfitab,3);
1218
1219             /* COULOMB ELECTROSTATICS */
1220             velec            = _mm_mul_pd(qq00,rinv00);
1221             felec            = _mm_mul_pd(velec,rinvsq00);
1222
1223             /* CUBIC SPLINE TABLE DISPERSION */
1224             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1225             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
1226             GMX_MM_TRANSPOSE2_PD(Y,F);
1227             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1228             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
1229             GMX_MM_TRANSPOSE2_PD(G,H);
1230             Heps             = _mm_mul_pd(vfeps,H);
1231             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1232             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1233             fvdw6            = _mm_mul_pd(c6_00,FF);
1234
1235             /* CUBIC SPLINE TABLE REPULSION */
1236             vfitab           = _mm_add_epi32(vfitab,ifour);
1237             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1238             F                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) );
1239             GMX_MM_TRANSPOSE2_PD(Y,F);
1240             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1241             H                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,1) +2);
1242             GMX_MM_TRANSPOSE2_PD(G,H);
1243             Heps             = _mm_mul_pd(vfeps,H);
1244             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1245             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1246             fvdw12           = _mm_mul_pd(c12_00,FF);
1247             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1248
1249             fscal            = _mm_add_pd(felec,fvdw);
1250
1251             /* Calculate temporary vectorial force */
1252             tx               = _mm_mul_pd(fscal,dx00);
1253             ty               = _mm_mul_pd(fscal,dy00);
1254             tz               = _mm_mul_pd(fscal,dz00);
1255
1256             /* Update vectorial force */
1257             fix0             = _mm_add_pd(fix0,tx);
1258             fiy0             = _mm_add_pd(fiy0,ty);
1259             fiz0             = _mm_add_pd(fiz0,tz);
1260
1261             fjx0             = _mm_add_pd(fjx0,tx);
1262             fjy0             = _mm_add_pd(fjy0,ty);
1263             fjz0             = _mm_add_pd(fjz0,tz);
1264
1265             /**************************
1266              * CALCULATE INTERACTIONS *
1267              **************************/
1268
1269             /* COULOMB ELECTROSTATICS */
1270             velec            = _mm_mul_pd(qq01,rinv01);
1271             felec            = _mm_mul_pd(velec,rinvsq01);
1272
1273             fscal            = felec;
1274
1275             /* Calculate temporary vectorial force */
1276             tx               = _mm_mul_pd(fscal,dx01);
1277             ty               = _mm_mul_pd(fscal,dy01);
1278             tz               = _mm_mul_pd(fscal,dz01);
1279
1280             /* Update vectorial force */
1281             fix0             = _mm_add_pd(fix0,tx);
1282             fiy0             = _mm_add_pd(fiy0,ty);
1283             fiz0             = _mm_add_pd(fiz0,tz);
1284
1285             fjx1             = _mm_add_pd(fjx1,tx);
1286             fjy1             = _mm_add_pd(fjy1,ty);
1287             fjz1             = _mm_add_pd(fjz1,tz);
1288
1289             /**************************
1290              * CALCULATE INTERACTIONS *
1291              **************************/
1292
1293             /* COULOMB ELECTROSTATICS */
1294             velec            = _mm_mul_pd(qq02,rinv02);
1295             felec            = _mm_mul_pd(velec,rinvsq02);
1296
1297             fscal            = felec;
1298
1299             /* Calculate temporary vectorial force */
1300             tx               = _mm_mul_pd(fscal,dx02);
1301             ty               = _mm_mul_pd(fscal,dy02);
1302             tz               = _mm_mul_pd(fscal,dz02);
1303
1304             /* Update vectorial force */
1305             fix0             = _mm_add_pd(fix0,tx);
1306             fiy0             = _mm_add_pd(fiy0,ty);
1307             fiz0             = _mm_add_pd(fiz0,tz);
1308
1309             fjx2             = _mm_add_pd(fjx2,tx);
1310             fjy2             = _mm_add_pd(fjy2,ty);
1311             fjz2             = _mm_add_pd(fjz2,tz);
1312
1313             /**************************
1314              * CALCULATE INTERACTIONS *
1315              **************************/
1316
1317             /* COULOMB ELECTROSTATICS */
1318             velec            = _mm_mul_pd(qq10,rinv10);
1319             felec            = _mm_mul_pd(velec,rinvsq10);
1320
1321             fscal            = felec;
1322
1323             /* Calculate temporary vectorial force */
1324             tx               = _mm_mul_pd(fscal,dx10);
1325             ty               = _mm_mul_pd(fscal,dy10);
1326             tz               = _mm_mul_pd(fscal,dz10);
1327
1328             /* Update vectorial force */
1329             fix1             = _mm_add_pd(fix1,tx);
1330             fiy1             = _mm_add_pd(fiy1,ty);
1331             fiz1             = _mm_add_pd(fiz1,tz);
1332
1333             fjx0             = _mm_add_pd(fjx0,tx);
1334             fjy0             = _mm_add_pd(fjy0,ty);
1335             fjz0             = _mm_add_pd(fjz0,tz);
1336
1337             /**************************
1338              * CALCULATE INTERACTIONS *
1339              **************************/
1340
1341             /* COULOMB ELECTROSTATICS */
1342             velec            = _mm_mul_pd(qq11,rinv11);
1343             felec            = _mm_mul_pd(velec,rinvsq11);
1344
1345             fscal            = felec;
1346
1347             /* Calculate temporary vectorial force */
1348             tx               = _mm_mul_pd(fscal,dx11);
1349             ty               = _mm_mul_pd(fscal,dy11);
1350             tz               = _mm_mul_pd(fscal,dz11);
1351
1352             /* Update vectorial force */
1353             fix1             = _mm_add_pd(fix1,tx);
1354             fiy1             = _mm_add_pd(fiy1,ty);
1355             fiz1             = _mm_add_pd(fiz1,tz);
1356
1357             fjx1             = _mm_add_pd(fjx1,tx);
1358             fjy1             = _mm_add_pd(fjy1,ty);
1359             fjz1             = _mm_add_pd(fjz1,tz);
1360
1361             /**************************
1362              * CALCULATE INTERACTIONS *
1363              **************************/
1364
1365             /* COULOMB ELECTROSTATICS */
1366             velec            = _mm_mul_pd(qq12,rinv12);
1367             felec            = _mm_mul_pd(velec,rinvsq12);
1368
1369             fscal            = felec;
1370
1371             /* Calculate temporary vectorial force */
1372             tx               = _mm_mul_pd(fscal,dx12);
1373             ty               = _mm_mul_pd(fscal,dy12);
1374             tz               = _mm_mul_pd(fscal,dz12);
1375
1376             /* Update vectorial force */
1377             fix1             = _mm_add_pd(fix1,tx);
1378             fiy1             = _mm_add_pd(fiy1,ty);
1379             fiz1             = _mm_add_pd(fiz1,tz);
1380
1381             fjx2             = _mm_add_pd(fjx2,tx);
1382             fjy2             = _mm_add_pd(fjy2,ty);
1383             fjz2             = _mm_add_pd(fjz2,tz);
1384
1385             /**************************
1386              * CALCULATE INTERACTIONS *
1387              **************************/
1388
1389             /* COULOMB ELECTROSTATICS */
1390             velec            = _mm_mul_pd(qq20,rinv20);
1391             felec            = _mm_mul_pd(velec,rinvsq20);
1392
1393             fscal            = felec;
1394
1395             /* Calculate temporary vectorial force */
1396             tx               = _mm_mul_pd(fscal,dx20);
1397             ty               = _mm_mul_pd(fscal,dy20);
1398             tz               = _mm_mul_pd(fscal,dz20);
1399
1400             /* Update vectorial force */
1401             fix2             = _mm_add_pd(fix2,tx);
1402             fiy2             = _mm_add_pd(fiy2,ty);
1403             fiz2             = _mm_add_pd(fiz2,tz);
1404
1405             fjx0             = _mm_add_pd(fjx0,tx);
1406             fjy0             = _mm_add_pd(fjy0,ty);
1407             fjz0             = _mm_add_pd(fjz0,tz);
1408
1409             /**************************
1410              * CALCULATE INTERACTIONS *
1411              **************************/
1412
1413             /* COULOMB ELECTROSTATICS */
1414             velec            = _mm_mul_pd(qq21,rinv21);
1415             felec            = _mm_mul_pd(velec,rinvsq21);
1416
1417             fscal            = felec;
1418
1419             /* Calculate temporary vectorial force */
1420             tx               = _mm_mul_pd(fscal,dx21);
1421             ty               = _mm_mul_pd(fscal,dy21);
1422             tz               = _mm_mul_pd(fscal,dz21);
1423
1424             /* Update vectorial force */
1425             fix2             = _mm_add_pd(fix2,tx);
1426             fiy2             = _mm_add_pd(fiy2,ty);
1427             fiz2             = _mm_add_pd(fiz2,tz);
1428
1429             fjx1             = _mm_add_pd(fjx1,tx);
1430             fjy1             = _mm_add_pd(fjy1,ty);
1431             fjz1             = _mm_add_pd(fjz1,tz);
1432
1433             /**************************
1434              * CALCULATE INTERACTIONS *
1435              **************************/
1436
1437             /* COULOMB ELECTROSTATICS */
1438             velec            = _mm_mul_pd(qq22,rinv22);
1439             felec            = _mm_mul_pd(velec,rinvsq22);
1440
1441             fscal            = felec;
1442
1443             /* Calculate temporary vectorial force */
1444             tx               = _mm_mul_pd(fscal,dx22);
1445             ty               = _mm_mul_pd(fscal,dy22);
1446             tz               = _mm_mul_pd(fscal,dz22);
1447
1448             /* Update vectorial force */
1449             fix2             = _mm_add_pd(fix2,tx);
1450             fiy2             = _mm_add_pd(fiy2,ty);
1451             fiz2             = _mm_add_pd(fiz2,tz);
1452
1453             fjx2             = _mm_add_pd(fjx2,tx);
1454             fjy2             = _mm_add_pd(fjy2,ty);
1455             fjz2             = _mm_add_pd(fjz2,tz);
1456
1457             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1458
1459             /* Inner loop uses 270 flops */
1460         }
1461
1462         if(jidx<j_index_end)
1463         {
1464
1465             jnrA             = jjnr[jidx];
1466             j_coord_offsetA  = DIM*jnrA;
1467
1468             /* load j atom coordinates */
1469             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1470                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1471
1472             /* Calculate displacement vector */
1473             dx00             = _mm_sub_pd(ix0,jx0);
1474             dy00             = _mm_sub_pd(iy0,jy0);
1475             dz00             = _mm_sub_pd(iz0,jz0);
1476             dx01             = _mm_sub_pd(ix0,jx1);
1477             dy01             = _mm_sub_pd(iy0,jy1);
1478             dz01             = _mm_sub_pd(iz0,jz1);
1479             dx02             = _mm_sub_pd(ix0,jx2);
1480             dy02             = _mm_sub_pd(iy0,jy2);
1481             dz02             = _mm_sub_pd(iz0,jz2);
1482             dx10             = _mm_sub_pd(ix1,jx0);
1483             dy10             = _mm_sub_pd(iy1,jy0);
1484             dz10             = _mm_sub_pd(iz1,jz0);
1485             dx11             = _mm_sub_pd(ix1,jx1);
1486             dy11             = _mm_sub_pd(iy1,jy1);
1487             dz11             = _mm_sub_pd(iz1,jz1);
1488             dx12             = _mm_sub_pd(ix1,jx2);
1489             dy12             = _mm_sub_pd(iy1,jy2);
1490             dz12             = _mm_sub_pd(iz1,jz2);
1491             dx20             = _mm_sub_pd(ix2,jx0);
1492             dy20             = _mm_sub_pd(iy2,jy0);
1493             dz20             = _mm_sub_pd(iz2,jz0);
1494             dx21             = _mm_sub_pd(ix2,jx1);
1495             dy21             = _mm_sub_pd(iy2,jy1);
1496             dz21             = _mm_sub_pd(iz2,jz1);
1497             dx22             = _mm_sub_pd(ix2,jx2);
1498             dy22             = _mm_sub_pd(iy2,jy2);
1499             dz22             = _mm_sub_pd(iz2,jz2);
1500
1501             /* Calculate squared distance and things based on it */
1502             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1503             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1504             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1505             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1506             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1507             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1508             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1509             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1510             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1511
1512             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1513             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1514             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1515             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1516             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1517             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1518             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1519             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1520             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1521
1522             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1523             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1524             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1525             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1526             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1527             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1528             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1529             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1530             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1531
1532             fjx0             = _mm_setzero_pd();
1533             fjy0             = _mm_setzero_pd();
1534             fjz0             = _mm_setzero_pd();
1535             fjx1             = _mm_setzero_pd();
1536             fjy1             = _mm_setzero_pd();
1537             fjz1             = _mm_setzero_pd();
1538             fjx2             = _mm_setzero_pd();
1539             fjy2             = _mm_setzero_pd();
1540             fjz2             = _mm_setzero_pd();
1541
1542             /**************************
1543              * CALCULATE INTERACTIONS *
1544              **************************/
1545
1546             r00              = _mm_mul_pd(rsq00,rinv00);
1547
1548             /* Calculate table index by multiplying r with table scale and truncate to integer */
1549             rt               = _mm_mul_pd(r00,vftabscale);
1550             vfitab           = _mm_cvttpd_epi32(rt);
1551             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
1552             vfitab           = _mm_slli_epi32(vfitab,3);
1553
1554             /* COULOMB ELECTROSTATICS */
1555             velec            = _mm_mul_pd(qq00,rinv00);
1556             felec            = _mm_mul_pd(velec,rinvsq00);
1557
1558             /* CUBIC SPLINE TABLE DISPERSION */
1559             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1560             F                = _mm_setzero_pd();
1561             GMX_MM_TRANSPOSE2_PD(Y,F);
1562             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1563             H                = _mm_setzero_pd();
1564             GMX_MM_TRANSPOSE2_PD(G,H);
1565             Heps             = _mm_mul_pd(vfeps,H);
1566             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1567             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1568             fvdw6            = _mm_mul_pd(c6_00,FF);
1569
1570             /* CUBIC SPLINE TABLE REPULSION */
1571             vfitab           = _mm_add_epi32(vfitab,ifour);
1572             Y                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) );
1573             F                = _mm_setzero_pd();
1574             GMX_MM_TRANSPOSE2_PD(Y,F);
1575             G                = _mm_load_pd( vftab + gmx_mm_extract_epi32(vfitab,0) +2);
1576             H                = _mm_setzero_pd();
1577             GMX_MM_TRANSPOSE2_PD(G,H);
1578             Heps             = _mm_mul_pd(vfeps,H);
1579             Fp               = _mm_add_pd(F,_mm_mul_pd(vfeps,_mm_add_pd(G,Heps)));
1580             FF               = _mm_add_pd(Fp,_mm_mul_pd(vfeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
1581             fvdw12           = _mm_mul_pd(c12_00,FF);
1582             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1583
1584             fscal            = _mm_add_pd(felec,fvdw);
1585
1586             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1587
1588             /* Calculate temporary vectorial force */
1589             tx               = _mm_mul_pd(fscal,dx00);
1590             ty               = _mm_mul_pd(fscal,dy00);
1591             tz               = _mm_mul_pd(fscal,dz00);
1592
1593             /* Update vectorial force */
1594             fix0             = _mm_add_pd(fix0,tx);
1595             fiy0             = _mm_add_pd(fiy0,ty);
1596             fiz0             = _mm_add_pd(fiz0,tz);
1597
1598             fjx0             = _mm_add_pd(fjx0,tx);
1599             fjy0             = _mm_add_pd(fjy0,ty);
1600             fjz0             = _mm_add_pd(fjz0,tz);
1601
1602             /**************************
1603              * CALCULATE INTERACTIONS *
1604              **************************/
1605
1606             /* COULOMB ELECTROSTATICS */
1607             velec            = _mm_mul_pd(qq01,rinv01);
1608             felec            = _mm_mul_pd(velec,rinvsq01);
1609
1610             fscal            = felec;
1611
1612             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1613
1614             /* Calculate temporary vectorial force */
1615             tx               = _mm_mul_pd(fscal,dx01);
1616             ty               = _mm_mul_pd(fscal,dy01);
1617             tz               = _mm_mul_pd(fscal,dz01);
1618
1619             /* Update vectorial force */
1620             fix0             = _mm_add_pd(fix0,tx);
1621             fiy0             = _mm_add_pd(fiy0,ty);
1622             fiz0             = _mm_add_pd(fiz0,tz);
1623
1624             fjx1             = _mm_add_pd(fjx1,tx);
1625             fjy1             = _mm_add_pd(fjy1,ty);
1626             fjz1             = _mm_add_pd(fjz1,tz);
1627
1628             /**************************
1629              * CALCULATE INTERACTIONS *
1630              **************************/
1631
1632             /* COULOMB ELECTROSTATICS */
1633             velec            = _mm_mul_pd(qq02,rinv02);
1634             felec            = _mm_mul_pd(velec,rinvsq02);
1635
1636             fscal            = felec;
1637
1638             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1639
1640             /* Calculate temporary vectorial force */
1641             tx               = _mm_mul_pd(fscal,dx02);
1642             ty               = _mm_mul_pd(fscal,dy02);
1643             tz               = _mm_mul_pd(fscal,dz02);
1644
1645             /* Update vectorial force */
1646             fix0             = _mm_add_pd(fix0,tx);
1647             fiy0             = _mm_add_pd(fiy0,ty);
1648             fiz0             = _mm_add_pd(fiz0,tz);
1649
1650             fjx2             = _mm_add_pd(fjx2,tx);
1651             fjy2             = _mm_add_pd(fjy2,ty);
1652             fjz2             = _mm_add_pd(fjz2,tz);
1653
1654             /**************************
1655              * CALCULATE INTERACTIONS *
1656              **************************/
1657
1658             /* COULOMB ELECTROSTATICS */
1659             velec            = _mm_mul_pd(qq10,rinv10);
1660             felec            = _mm_mul_pd(velec,rinvsq10);
1661
1662             fscal            = felec;
1663
1664             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1665
1666             /* Calculate temporary vectorial force */
1667             tx               = _mm_mul_pd(fscal,dx10);
1668             ty               = _mm_mul_pd(fscal,dy10);
1669             tz               = _mm_mul_pd(fscal,dz10);
1670
1671             /* Update vectorial force */
1672             fix1             = _mm_add_pd(fix1,tx);
1673             fiy1             = _mm_add_pd(fiy1,ty);
1674             fiz1             = _mm_add_pd(fiz1,tz);
1675
1676             fjx0             = _mm_add_pd(fjx0,tx);
1677             fjy0             = _mm_add_pd(fjy0,ty);
1678             fjz0             = _mm_add_pd(fjz0,tz);
1679
1680             /**************************
1681              * CALCULATE INTERACTIONS *
1682              **************************/
1683
1684             /* COULOMB ELECTROSTATICS */
1685             velec            = _mm_mul_pd(qq11,rinv11);
1686             felec            = _mm_mul_pd(velec,rinvsq11);
1687
1688             fscal            = felec;
1689
1690             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1691
1692             /* Calculate temporary vectorial force */
1693             tx               = _mm_mul_pd(fscal,dx11);
1694             ty               = _mm_mul_pd(fscal,dy11);
1695             tz               = _mm_mul_pd(fscal,dz11);
1696
1697             /* Update vectorial force */
1698             fix1             = _mm_add_pd(fix1,tx);
1699             fiy1             = _mm_add_pd(fiy1,ty);
1700             fiz1             = _mm_add_pd(fiz1,tz);
1701
1702             fjx1             = _mm_add_pd(fjx1,tx);
1703             fjy1             = _mm_add_pd(fjy1,ty);
1704             fjz1             = _mm_add_pd(fjz1,tz);
1705
1706             /**************************
1707              * CALCULATE INTERACTIONS *
1708              **************************/
1709
1710             /* COULOMB ELECTROSTATICS */
1711             velec            = _mm_mul_pd(qq12,rinv12);
1712             felec            = _mm_mul_pd(velec,rinvsq12);
1713
1714             fscal            = felec;
1715
1716             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1717
1718             /* Calculate temporary vectorial force */
1719             tx               = _mm_mul_pd(fscal,dx12);
1720             ty               = _mm_mul_pd(fscal,dy12);
1721             tz               = _mm_mul_pd(fscal,dz12);
1722
1723             /* Update vectorial force */
1724             fix1             = _mm_add_pd(fix1,tx);
1725             fiy1             = _mm_add_pd(fiy1,ty);
1726             fiz1             = _mm_add_pd(fiz1,tz);
1727
1728             fjx2             = _mm_add_pd(fjx2,tx);
1729             fjy2             = _mm_add_pd(fjy2,ty);
1730             fjz2             = _mm_add_pd(fjz2,tz);
1731
1732             /**************************
1733              * CALCULATE INTERACTIONS *
1734              **************************/
1735
1736             /* COULOMB ELECTROSTATICS */
1737             velec            = _mm_mul_pd(qq20,rinv20);
1738             felec            = _mm_mul_pd(velec,rinvsq20);
1739
1740             fscal            = felec;
1741
1742             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1743
1744             /* Calculate temporary vectorial force */
1745             tx               = _mm_mul_pd(fscal,dx20);
1746             ty               = _mm_mul_pd(fscal,dy20);
1747             tz               = _mm_mul_pd(fscal,dz20);
1748
1749             /* Update vectorial force */
1750             fix2             = _mm_add_pd(fix2,tx);
1751             fiy2             = _mm_add_pd(fiy2,ty);
1752             fiz2             = _mm_add_pd(fiz2,tz);
1753
1754             fjx0             = _mm_add_pd(fjx0,tx);
1755             fjy0             = _mm_add_pd(fjy0,ty);
1756             fjz0             = _mm_add_pd(fjz0,tz);
1757
1758             /**************************
1759              * CALCULATE INTERACTIONS *
1760              **************************/
1761
1762             /* COULOMB ELECTROSTATICS */
1763             velec            = _mm_mul_pd(qq21,rinv21);
1764             felec            = _mm_mul_pd(velec,rinvsq21);
1765
1766             fscal            = felec;
1767
1768             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1769
1770             /* Calculate temporary vectorial force */
1771             tx               = _mm_mul_pd(fscal,dx21);
1772             ty               = _mm_mul_pd(fscal,dy21);
1773             tz               = _mm_mul_pd(fscal,dz21);
1774
1775             /* Update vectorial force */
1776             fix2             = _mm_add_pd(fix2,tx);
1777             fiy2             = _mm_add_pd(fiy2,ty);
1778             fiz2             = _mm_add_pd(fiz2,tz);
1779
1780             fjx1             = _mm_add_pd(fjx1,tx);
1781             fjy1             = _mm_add_pd(fjy1,ty);
1782             fjz1             = _mm_add_pd(fjz1,tz);
1783
1784             /**************************
1785              * CALCULATE INTERACTIONS *
1786              **************************/
1787
1788             /* COULOMB ELECTROSTATICS */
1789             velec            = _mm_mul_pd(qq22,rinv22);
1790             felec            = _mm_mul_pd(velec,rinvsq22);
1791
1792             fscal            = felec;
1793
1794             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1795
1796             /* Calculate temporary vectorial force */
1797             tx               = _mm_mul_pd(fscal,dx22);
1798             ty               = _mm_mul_pd(fscal,dy22);
1799             tz               = _mm_mul_pd(fscal,dz22);
1800
1801             /* Update vectorial force */
1802             fix2             = _mm_add_pd(fix2,tx);
1803             fiy2             = _mm_add_pd(fiy2,ty);
1804             fiz2             = _mm_add_pd(fiz2,tz);
1805
1806             fjx2             = _mm_add_pd(fjx2,tx);
1807             fjy2             = _mm_add_pd(fjy2,ty);
1808             fjz2             = _mm_add_pd(fjz2,tz);
1809
1810             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1811
1812             /* Inner loop uses 270 flops */
1813         }
1814
1815         /* End of innermost loop */
1816
1817         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1818                                               f+i_coord_offset,fshift+i_shift_offset);
1819
1820         /* Increment number of inner iterations */
1821         inneriter                  += j_index_end - j_index_start;
1822
1823         /* Outer loop uses 18 flops */
1824     }
1825
1826     /* Increment number of outer iterations */
1827     outeriter        += nri;
1828
1829     /* Update outer/inner flops */
1830
1831     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*270);
1832 }