src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecCoul_VdwCSTab_GeomW4W4_avx_128_fma_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_double kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_double.h"
  48 #include "kernelutil_x86_avx_128_fma_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwCSTab_GeomW4W4_VF_avx_128_fma_double
  52  * Electrostatics interaction: Coulomb
  53  * VdW interaction:            CubicSplineTable
  54  * Geometry:                   Water4-Water4
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecCoul_VdwCSTab_GeomW4W4_VF_avx_128_fma_double
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB;
  75     int              j_coord_offsetA,j_coord_offsetB;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  80     int              vdwioffset0;
  81     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  82     int              vdwioffset1;
  83     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  84     int              vdwioffset2;
  85     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  86     int              vdwioffset3;
  87     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  88     int              vdwjidx0A,vdwjidx0B;
  89     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     int              vdwjidx1A,vdwjidx1B;
  91     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  92     int              vdwjidx2A,vdwjidx2B;
  93     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  94     int              vdwjidx3A,vdwjidx3B;
  95     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  96     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  97     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  98     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  99     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 100     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 101     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 102     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 103     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 104     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 105     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 106     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 107     real             *charge;
 108     int              nvdwtype;
 109     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 110     int              *vdwtype;
 111     real             *vdwparam;
 112     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 113     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 114     __m128i          vfitab;
 115     __m128i          ifour       = _mm_set1_epi32(4);
 116     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
 117     real             *vftab;
 118     __m128d          dummy_mask,cutoff_mask;
 119     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 120     __m128d          one     = _mm_set1_pd(1.0);
 121     __m128d          two     = _mm_set1_pd(2.0);
 122     x                = xx[0];
 123     f                = ff[0];
 124
 125     nri              = nlist->nri;
 126     iinr             = nlist->iinr;
 127     jindex           = nlist->jindex;
 128     jjnr             = nlist->jjnr;
 129     shiftidx         = nlist->shift;
 130     gid              = nlist->gid;
 131     shiftvec         = fr->shift_vec[0];
 132     fshift           = fr->fshift[0];
 133     facel            = _mm_set1_pd(fr->epsfac);
 134     charge           = mdatoms->chargeA;
 135     nvdwtype         = fr->ntype;
 136     vdwparam         = fr->nbfp;
 137     vdwtype          = mdatoms->typeA;
 138
 139     vftab            = kernel_data->table_vdw->data;
 140     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
 141
 142     /* Setup water-specific parameters */
 143     inr              = nlist->iinr[0];
 144     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 145     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 146     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 147     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 148
 149     jq1              = _mm_set1_pd(charge[inr+1]);
 150     jq2              = _mm_set1_pd(charge[inr+2]);
 151     jq3              = _mm_set1_pd(charge[inr+3]);
 152     vdwjidx0A        = 2*vdwtype[inr+0];
 153     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 154     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 155     qq11             = _mm_mul_pd(iq1,jq1);
 156     qq12             = _mm_mul_pd(iq1,jq2);
 157     qq13             = _mm_mul_pd(iq1,jq3);
 158     qq21             = _mm_mul_pd(iq2,jq1);
 159     qq22             = _mm_mul_pd(iq2,jq2);
 160     qq23             = _mm_mul_pd(iq2,jq3);
 161     qq31             = _mm_mul_pd(iq3,jq1);
 162     qq32             = _mm_mul_pd(iq3,jq2);
 163     qq33             = _mm_mul_pd(iq3,jq3);
 164
 165     /* Avoid stupid compiler warnings */
 166     jnrA = jnrB = 0;
 167     j_coord_offsetA = 0;
 168     j_coord_offsetB = 0;
 169
 170     outeriter        = 0;
 171     inneriter        = 0;
 172
 173     /* Start outer loop over neighborlists */
 174     for(iidx=0; iidx<nri; iidx++)
 175     {
 176         /* Load shift vector for this list */
 177         i_shift_offset   = DIM*shiftidx[iidx];
 178
 179         /* Load limits for loop over neighbors */
 180         j_index_start    = jindex[iidx];
 181         j_index_end      = jindex[iidx+1];
 182
 183         /* Get outer coordinate index */
 184         inr              = iinr[iidx];
 185         i_coord_offset   = DIM*inr;
 186
 187         /* Load i particle coords and add shift vector */
 188         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 189                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 190
 191         fix0             = _mm_setzero_pd();
 192         fiy0             = _mm_setzero_pd();
 193         fiz0             = _mm_setzero_pd();
 194         fix1             = _mm_setzero_pd();
 195         fiy1             = _mm_setzero_pd();
 196         fiz1             = _mm_setzero_pd();
 197         fix2             = _mm_setzero_pd();
 198         fiy2             = _mm_setzero_pd();
 199         fiz2             = _mm_setzero_pd();
 200         fix3             = _mm_setzero_pd();
 201         fiy3             = _mm_setzero_pd();
 202         fiz3             = _mm_setzero_pd();
 203
 204         /* Reset potential sums */
 205         velecsum         = _mm_setzero_pd();
 206         vvdwsum          = _mm_setzero_pd();
 207
 208         /* Start inner kernel loop */
 209         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 210         {
 211
 212             /* Get j neighbor index, and coordinate index */
 213             jnrA             = jjnr[jidx];
 214             jnrB             = jjnr[jidx+1];
 215             j_coord_offsetA  = DIM*jnrA;
 216             j_coord_offsetB  = DIM*jnrB;
 217
 218             /* load j atom coordinates */
 219             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 220                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 221                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 222
 223             /* Calculate displacement vector */
 224             dx00             = _mm_sub_pd(ix0,jx0);
 225             dy00             = _mm_sub_pd(iy0,jy0);
 226             dz00             = _mm_sub_pd(iz0,jz0);
 227             dx11             = _mm_sub_pd(ix1,jx1);
 228             dy11             = _mm_sub_pd(iy1,jy1);
 229             dz11             = _mm_sub_pd(iz1,jz1);
 230             dx12             = _mm_sub_pd(ix1,jx2);
 231             dy12             = _mm_sub_pd(iy1,jy2);
 232             dz12             = _mm_sub_pd(iz1,jz2);
 233             dx13             = _mm_sub_pd(ix1,jx3);
 234             dy13             = _mm_sub_pd(iy1,jy3);
 235             dz13             = _mm_sub_pd(iz1,jz3);
 236             dx21             = _mm_sub_pd(ix2,jx1);
 237             dy21             = _mm_sub_pd(iy2,jy1);
 238             dz21             = _mm_sub_pd(iz2,jz1);
 239             dx22             = _mm_sub_pd(ix2,jx2);
 240             dy22             = _mm_sub_pd(iy2,jy2);
 241             dz22             = _mm_sub_pd(iz2,jz2);
 242             dx23             = _mm_sub_pd(ix2,jx3);
 243             dy23             = _mm_sub_pd(iy2,jy3);
 244             dz23             = _mm_sub_pd(iz2,jz3);
 245             dx31             = _mm_sub_pd(ix3,jx1);
 246             dy31             = _mm_sub_pd(iy3,jy1);
 247             dz31             = _mm_sub_pd(iz3,jz1);
 248             dx32             = _mm_sub_pd(ix3,jx2);
 249             dy32             = _mm_sub_pd(iy3,jy2);
 250             dz32             = _mm_sub_pd(iz3,jz2);
 251             dx33             = _mm_sub_pd(ix3,jx3);
 252             dy33             = _mm_sub_pd(iy3,jy3);
 253             dz33             = _mm_sub_pd(iz3,jz3);
 254
 255             /* Calculate squared distance and things based on it */
 256             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 257             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 258             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 259             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 260             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 261             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 262             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 263             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 264             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 265             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 266
 267             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 268             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 269             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 270             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 271             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 272             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 273             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 274             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 275             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 276             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 277
 278             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 279             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 280             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 281             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 282             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 283             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 284             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 285             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 286             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 287
 288             fjx0             = _mm_setzero_pd();
 289             fjy0             = _mm_setzero_pd();
 290             fjz0             = _mm_setzero_pd();
 291             fjx1             = _mm_setzero_pd();
 292             fjy1             = _mm_setzero_pd();
 293             fjz1             = _mm_setzero_pd();
 294             fjx2             = _mm_setzero_pd();
 295             fjy2             = _mm_setzero_pd();
 296             fjz2             = _mm_setzero_pd();
 297             fjx3             = _mm_setzero_pd();
 298             fjy3             = _mm_setzero_pd();
 299             fjz3             = _mm_setzero_pd();
 300
 301             /**************************
 302              * CALCULATE INTERACTIONS *
 303              **************************/
 304
 305             r00              = _mm_mul_pd(rsq00,rinv00);
 306
 307             /* Calculate table index by multiplying r with table scale and truncate to integer */
 308             rt               = _mm_mul_pd(r00,vftabscale);
 309             vfitab           = _mm_cvttpd_epi32(rt);
 310 #ifdef __XOP__
 311             vfeps            = _mm_frcz_pd(rt);
 312 #else
 313             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 314 #endif
 315             twovfeps         = _mm_add_pd(vfeps,vfeps);
 316             vfitab           = _mm_slli_epi32(vfitab,3);
 317
 318             /* CUBIC SPLINE TABLE DISPERSION */
 319             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 320             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 321             GMX_MM_TRANSPOSE2_PD(Y,F);
 322             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 323             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
 324             GMX_MM_TRANSPOSE2_PD(G,H);
 325             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 326             VV               = _mm_macc_pd(vfeps,Fp,Y);
 327             vvdw6            = _mm_mul_pd(c6_00,VV);
 328             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 329             fvdw6            = _mm_mul_pd(c6_00,FF);
 330
 331             /* CUBIC SPLINE TABLE REPULSION */
 332             vfitab           = _mm_add_epi32(vfitab,ifour);
 333             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 334             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 335             GMX_MM_TRANSPOSE2_PD(Y,F);
 336             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 337             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
 338             GMX_MM_TRANSPOSE2_PD(G,H);
 339             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 340             VV               = _mm_macc_pd(vfeps,Fp,Y);
 341             vvdw12           = _mm_mul_pd(c12_00,VV);
 342             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 343             fvdw12           = _mm_mul_pd(c12_00,FF);
 344             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 345             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 346
 347             /* Update potential sum for this i atom from the interaction with this j atom. */
 348             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 349
 350             fscal            = fvdw;
 351
 352             /* Update vectorial force */
 353             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 354             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 355             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 356
 357             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 358             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 359             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 360
 361             /**************************
 362              * CALCULATE INTERACTIONS *
 363              **************************/
 364
 365             /* COULOMB ELECTROSTATICS */
 366             velec            = _mm_mul_pd(qq11,rinv11);
 367             felec            = _mm_mul_pd(velec,rinvsq11);
 368
 369             /* Update potential sum for this i atom from the interaction with this j atom. */
 370             velecsum         = _mm_add_pd(velecsum,velec);
 371
 372             fscal            = felec;
 373
 374             /* Update vectorial force */
 375             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 376             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 377             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 378
 379             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 380             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 381             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 382
 383             /**************************
 384              * CALCULATE INTERACTIONS *
 385              **************************/
 386
 387             /* COULOMB ELECTROSTATICS */
 388             velec            = _mm_mul_pd(qq12,rinv12);
 389             felec            = _mm_mul_pd(velec,rinvsq12);
 390
 391             /* Update potential sum for this i atom from the interaction with this j atom. */
 392             velecsum         = _mm_add_pd(velecsum,velec);
 393
 394             fscal            = felec;
 395
 396             /* Update vectorial force */
 397             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 398             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 399             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 400
 401             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 402             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 403             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 404
 405             /**************************
 406              * CALCULATE INTERACTIONS *
 407              **************************/
 408
 409             /* COULOMB ELECTROSTATICS */
 410             velec            = _mm_mul_pd(qq13,rinv13);
 411             felec            = _mm_mul_pd(velec,rinvsq13);
 412
 413             /* Update potential sum for this i atom from the interaction with this j atom. */
 414             velecsum         = _mm_add_pd(velecsum,velec);
 415
 416             fscal            = felec;
 417
 418             /* Update vectorial force */
 419             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 420             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 421             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 422
 423             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 424             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 425             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 426
 427             /**************************
 428              * CALCULATE INTERACTIONS *
 429              **************************/
 430
 431             /* COULOMB ELECTROSTATICS */
 432             velec            = _mm_mul_pd(qq21,rinv21);
 433             felec            = _mm_mul_pd(velec,rinvsq21);
 434
 435             /* Update potential sum for this i atom from the interaction with this j atom. */
 436             velecsum         = _mm_add_pd(velecsum,velec);
 437
 438             fscal            = felec;
 439
 440             /* Update vectorial force */
 441             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 442             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 443             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 444
 445             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 446             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 447             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 448
 449             /**************************
 450              * CALCULATE INTERACTIONS *
 451              **************************/
 452
 453             /* COULOMB ELECTROSTATICS */
 454             velec            = _mm_mul_pd(qq22,rinv22);
 455             felec            = _mm_mul_pd(velec,rinvsq22);
 456
 457             /* Update potential sum for this i atom from the interaction with this j atom. */
 458             velecsum         = _mm_add_pd(velecsum,velec);
 459
 460             fscal            = felec;
 461
 462             /* Update vectorial force */
 463             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 464             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 465             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 466
 467             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 468             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 469             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 470
 471             /**************************
 472              * CALCULATE INTERACTIONS *
 473              **************************/
 474
 475             /* COULOMB ELECTROSTATICS */
 476             velec            = _mm_mul_pd(qq23,rinv23);
 477             felec            = _mm_mul_pd(velec,rinvsq23);
 478
 479             /* Update potential sum for this i atom from the interaction with this j atom. */
 480             velecsum         = _mm_add_pd(velecsum,velec);
 481
 482             fscal            = felec;
 483
 484             /* Update vectorial force */
 485             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 486             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 487             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 488
 489             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 490             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 491             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 492
 493             /**************************
 494              * CALCULATE INTERACTIONS *
 495              **************************/
 496
 497             /* COULOMB ELECTROSTATICS */
 498             velec            = _mm_mul_pd(qq31,rinv31);
 499             felec            = _mm_mul_pd(velec,rinvsq31);
 500
 501             /* Update potential sum for this i atom from the interaction with this j atom. */
 502             velecsum         = _mm_add_pd(velecsum,velec);
 503
 504             fscal            = felec;
 505
 506             /* Update vectorial force */
 507             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 508             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 509             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 510
 511             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 512             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 513             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 514
 515             /**************************
 516              * CALCULATE INTERACTIONS *
 517              **************************/
 518
 519             /* COULOMB ELECTROSTATICS */
 520             velec            = _mm_mul_pd(qq32,rinv32);
 521             felec            = _mm_mul_pd(velec,rinvsq32);
 522
 523             /* Update potential sum for this i atom from the interaction with this j atom. */
 524             velecsum         = _mm_add_pd(velecsum,velec);
 525
 526             fscal            = felec;
 527
 528             /* Update vectorial force */
 529             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 530             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 531             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 532
 533             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 534             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 535             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 536
 537             /**************************
 538              * CALCULATE INTERACTIONS *
 539              **************************/
 540
 541             /* COULOMB ELECTROSTATICS */
 542             velec            = _mm_mul_pd(qq33,rinv33);
 543             felec            = _mm_mul_pd(velec,rinvsq33);
 544
 545             /* Update potential sum for this i atom from the interaction with this j atom. */
 546             velecsum         = _mm_add_pd(velecsum,velec);
 547
 548             fscal            = felec;
 549
 550             /* Update vectorial force */
 551             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 552             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 553             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 554
 555             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 556             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 557             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 558
 559             gmx_mm_decrement_4rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 560
 561             /* Inner loop uses 341 flops */
 562         }
 563
 564         if(jidx<j_index_end)
 565         {
 566
 567             jnrA             = jjnr[jidx];
 568             j_coord_offsetA  = DIM*jnrA;
 569
 570             /* load j atom coordinates */
 571             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 572                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 573                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 574
 575             /* Calculate displacement vector */
 576             dx00             = _mm_sub_pd(ix0,jx0);
 577             dy00             = _mm_sub_pd(iy0,jy0);
 578             dz00             = _mm_sub_pd(iz0,jz0);
 579             dx11             = _mm_sub_pd(ix1,jx1);
 580             dy11             = _mm_sub_pd(iy1,jy1);
 581             dz11             = _mm_sub_pd(iz1,jz1);
 582             dx12             = _mm_sub_pd(ix1,jx2);
 583             dy12             = _mm_sub_pd(iy1,jy2);
 584             dz12             = _mm_sub_pd(iz1,jz2);
 585             dx13             = _mm_sub_pd(ix1,jx3);
 586             dy13             = _mm_sub_pd(iy1,jy3);
 587             dz13             = _mm_sub_pd(iz1,jz3);
 588             dx21             = _mm_sub_pd(ix2,jx1);
 589             dy21             = _mm_sub_pd(iy2,jy1);
 590             dz21             = _mm_sub_pd(iz2,jz1);
 591             dx22             = _mm_sub_pd(ix2,jx2);
 592             dy22             = _mm_sub_pd(iy2,jy2);
 593             dz22             = _mm_sub_pd(iz2,jz2);
 594             dx23             = _mm_sub_pd(ix2,jx3);
 595             dy23             = _mm_sub_pd(iy2,jy3);
 596             dz23             = _mm_sub_pd(iz2,jz3);
 597             dx31             = _mm_sub_pd(ix3,jx1);
 598             dy31             = _mm_sub_pd(iy3,jy1);
 599             dz31             = _mm_sub_pd(iz3,jz1);
 600             dx32             = _mm_sub_pd(ix3,jx2);
 601             dy32             = _mm_sub_pd(iy3,jy2);
 602             dz32             = _mm_sub_pd(iz3,jz2);
 603             dx33             = _mm_sub_pd(ix3,jx3);
 604             dy33             = _mm_sub_pd(iy3,jy3);
 605             dz33             = _mm_sub_pd(iz3,jz3);
 606
 607             /* Calculate squared distance and things based on it */
 608             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 609             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 610             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 611             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 612             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 613             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 614             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 615             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 616             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 617             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 618
 619             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 620             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 621             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 622             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 623             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 624             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 625             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 626             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 627             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 628             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 629
 630             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 631             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 632             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 633             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 634             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 635             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 636             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 637             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 638             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 639
 640             fjx0             = _mm_setzero_pd();
 641             fjy0             = _mm_setzero_pd();
 642             fjz0             = _mm_setzero_pd();
 643             fjx1             = _mm_setzero_pd();
 644             fjy1             = _mm_setzero_pd();
 645             fjz1             = _mm_setzero_pd();
 646             fjx2             = _mm_setzero_pd();
 647             fjy2             = _mm_setzero_pd();
 648             fjz2             = _mm_setzero_pd();
 649             fjx3             = _mm_setzero_pd();
 650             fjy3             = _mm_setzero_pd();
 651             fjz3             = _mm_setzero_pd();
 652
 653             /**************************
 654              * CALCULATE INTERACTIONS *
 655              **************************/
 656
 657             r00              = _mm_mul_pd(rsq00,rinv00);
 658
 659             /* Calculate table index by multiplying r with table scale and truncate to integer */
 660             rt               = _mm_mul_pd(r00,vftabscale);
 661             vfitab           = _mm_cvttpd_epi32(rt);
 662 #ifdef __XOP__
 663             vfeps            = _mm_frcz_pd(rt);
 664 #else
 665             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 666 #endif
 667             twovfeps         = _mm_add_pd(vfeps,vfeps);
 668             vfitab           = _mm_slli_epi32(vfitab,3);
 669
 670             /* CUBIC SPLINE TABLE DISPERSION */
 671             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 672             F                = _mm_setzero_pd();
 673             GMX_MM_TRANSPOSE2_PD(Y,F);
 674             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 675             H                = _mm_setzero_pd();
 676             GMX_MM_TRANSPOSE2_PD(G,H);
 677             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 678             VV               = _mm_macc_pd(vfeps,Fp,Y);
 679             vvdw6            = _mm_mul_pd(c6_00,VV);
 680             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 681             fvdw6            = _mm_mul_pd(c6_00,FF);
 682
 683             /* CUBIC SPLINE TABLE REPULSION */
 684             vfitab           = _mm_add_epi32(vfitab,ifour);
 685             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 686             F                = _mm_setzero_pd();
 687             GMX_MM_TRANSPOSE2_PD(Y,F);
 688             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 689             H                = _mm_setzero_pd();
 690             GMX_MM_TRANSPOSE2_PD(G,H);
 691             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 692             VV               = _mm_macc_pd(vfeps,Fp,Y);
 693             vvdw12           = _mm_mul_pd(c12_00,VV);
 694             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 695             fvdw12           = _mm_mul_pd(c12_00,FF);
 696             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 697             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 698
 699             /* Update potential sum for this i atom from the interaction with this j atom. */
 700             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 701             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 702
 703             fscal            = fvdw;
 704
 705             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 706
 707             /* Update vectorial force */
 708             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 709             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 710             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 711
 712             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 713             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 714             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 715
 716             /**************************
 717              * CALCULATE INTERACTIONS *
 718              **************************/
 719
 720             /* COULOMB ELECTROSTATICS */
 721             velec            = _mm_mul_pd(qq11,rinv11);
 722             felec            = _mm_mul_pd(velec,rinvsq11);
 723
 724             /* Update potential sum for this i atom from the interaction with this j atom. */
 725             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 726             velecsum         = _mm_add_pd(velecsum,velec);
 727
 728             fscal            = felec;
 729
 730             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 731
 732             /* Update vectorial force */
 733             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 734             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 735             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 736
 737             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 738             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 739             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 740
 741             /**************************
 742              * CALCULATE INTERACTIONS *
 743              **************************/
 744
 745             /* COULOMB ELECTROSTATICS */
 746             velec            = _mm_mul_pd(qq12,rinv12);
 747             felec            = _mm_mul_pd(velec,rinvsq12);
 748
 749             /* Update potential sum for this i atom from the interaction with this j atom. */
 750             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 751             velecsum         = _mm_add_pd(velecsum,velec);
 752
 753             fscal            = felec;
 754
 755             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 756
 757             /* Update vectorial force */
 758             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 759             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 760             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 761
 762             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 763             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 764             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 765
 766             /**************************
 767              * CALCULATE INTERACTIONS *
 768              **************************/
 769
 770             /* COULOMB ELECTROSTATICS */
 771             velec            = _mm_mul_pd(qq13,rinv13);
 772             felec            = _mm_mul_pd(velec,rinvsq13);
 773
 774             /* Update potential sum for this i atom from the interaction with this j atom. */
 775             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 776             velecsum         = _mm_add_pd(velecsum,velec);
 777
 778             fscal            = felec;
 779
 780             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 781
 782             /* Update vectorial force */
 783             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 784             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 785             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 786
 787             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 788             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 789             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 790
 791             /**************************
 792              * CALCULATE INTERACTIONS *
 793              **************************/
 794
 795             /* COULOMB ELECTROSTATICS */
 796             velec            = _mm_mul_pd(qq21,rinv21);
 797             felec            = _mm_mul_pd(velec,rinvsq21);
 798
 799             /* Update potential sum for this i atom from the interaction with this j atom. */
 800             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 801             velecsum         = _mm_add_pd(velecsum,velec);
 802
 803             fscal            = felec;
 804
 805             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 806
 807             /* Update vectorial force */
 808             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 809             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 810             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 811
 812             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 813             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 814             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 815
 816             /**************************
 817              * CALCULATE INTERACTIONS *
 818              **************************/
 819
 820             /* COULOMB ELECTROSTATICS */
 821             velec            = _mm_mul_pd(qq22,rinv22);
 822             felec            = _mm_mul_pd(velec,rinvsq22);
 823
 824             /* Update potential sum for this i atom from the interaction with this j atom. */
 825             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 826             velecsum         = _mm_add_pd(velecsum,velec);
 827
 828             fscal            = felec;
 829
 830             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 831
 832             /* Update vectorial force */
 833             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 834             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 835             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 836
 837             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 838             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 839             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 840
 841             /**************************
 842              * CALCULATE INTERACTIONS *
 843              **************************/
 844
 845             /* COULOMB ELECTROSTATICS */
 846             velec            = _mm_mul_pd(qq23,rinv23);
 847             felec            = _mm_mul_pd(velec,rinvsq23);
 848
 849             /* Update potential sum for this i atom from the interaction with this j atom. */
 850             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 851             velecsum         = _mm_add_pd(velecsum,velec);
 852
 853             fscal            = felec;
 854
 855             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 856
 857             /* Update vectorial force */
 858             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 859             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 860             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 861
 862             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 863             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 864             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 865
 866             /**************************
 867              * CALCULATE INTERACTIONS *
 868              **************************/
 869
 870             /* COULOMB ELECTROSTATICS */
 871             velec            = _mm_mul_pd(qq31,rinv31);
 872             felec            = _mm_mul_pd(velec,rinvsq31);
 873
 874             /* Update potential sum for this i atom from the interaction with this j atom. */
 875             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 876             velecsum         = _mm_add_pd(velecsum,velec);
 877
 878             fscal            = felec;
 879
 880             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 881
 882             /* Update vectorial force */
 883             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 884             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 885             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 886
 887             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 888             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 889             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 890
 891             /**************************
 892              * CALCULATE INTERACTIONS *
 893              **************************/
 894
 895             /* COULOMB ELECTROSTATICS */
 896             velec            = _mm_mul_pd(qq32,rinv32);
 897             felec            = _mm_mul_pd(velec,rinvsq32);
 898
 899             /* Update potential sum for this i atom from the interaction with this j atom. */
 900             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 901             velecsum         = _mm_add_pd(velecsum,velec);
 902
 903             fscal            = felec;
 904
 905             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 906
 907             /* Update vectorial force */
 908             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 909             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 910             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 911
 912             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 913             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 914             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 915
 916             /**************************
 917              * CALCULATE INTERACTIONS *
 918              **************************/
 919
 920             /* COULOMB ELECTROSTATICS */
 921             velec            = _mm_mul_pd(qq33,rinv33);
 922             felec            = _mm_mul_pd(velec,rinvsq33);
 923
 924             /* Update potential sum for this i atom from the interaction with this j atom. */
 925             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 926             velecsum         = _mm_add_pd(velecsum,velec);
 927
 928             fscal            = felec;
 929
 930             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 931
 932             /* Update vectorial force */
 933             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 934             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 935             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 936
 937             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 938             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 939             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 940
 941             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 942
 943             /* Inner loop uses 341 flops */
 944         }
 945
 946         /* End of innermost loop */
 947
 948         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 949                                               f+i_coord_offset,fshift+i_shift_offset);
 950
 951         ggid                        = gid[iidx];
 952         /* Update potential energies */
 953         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 954         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 955
 956         /* Increment number of inner iterations */
 957         inneriter                  += j_index_end - j_index_start;
 958
 959         /* Outer loop uses 26 flops */
 960     }
 961
 962     /* Increment number of outer iterations */
 963     outeriter        += nri;
 964
 965     /* Update outer/inner flops */
 966
 967     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*341);
 968 }
 969 /*
 970  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwCSTab_GeomW4W4_F_avx_128_fma_double
 971  * Electrostatics interaction: Coulomb
 972  * VdW interaction:            CubicSplineTable
 973  * Geometry:                   Water4-Water4
 974  * Calculate force/pot:        Force
 975  */
 976 void
 977 nb_kernel_ElecCoul_VdwCSTab_GeomW4W4_F_avx_128_fma_double
 978                     (t_nblist                    * gmx_restrict       nlist,
 979                      rvec                        * gmx_restrict          xx,
 980                      rvec                        * gmx_restrict          ff,
 981                      t_forcerec                  * gmx_restrict          fr,
 982                      t_mdatoms                   * gmx_restrict     mdatoms,
 983                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 984                      t_nrnb                      * gmx_restrict        nrnb)
 985 {
 986     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 987      * just 0 for non-waters.
 988      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 989      * jnr indices corresponding to data put in the four positions in the SIMD register.
 990      */
 991     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 992     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 993     int              jnrA,jnrB;
 994     int              j_coord_offsetA,j_coord_offsetB;
 995     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 996     real             rcutoff_scalar;
 997     real             *shiftvec,*fshift,*x,*f;
 998     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 999     int              vdwioffset0;
1000     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1001     int              vdwioffset1;
1002     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1003     int              vdwioffset2;
1004     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1005     int              vdwioffset3;
1006     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1007     int              vdwjidx0A,vdwjidx0B;
1008     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1009     int              vdwjidx1A,vdwjidx1B;
1010     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1011     int              vdwjidx2A,vdwjidx2B;
1012     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1013     int              vdwjidx3A,vdwjidx3B;
1014     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1015     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1016     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1017     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1018     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1019     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1020     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1021     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1022     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1023     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1024     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1025     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1026     real             *charge;
1027     int              nvdwtype;
1028     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1029     int              *vdwtype;
1030     real             *vdwparam;
1031     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1032     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1033     __m128i          vfitab;
1034     __m128i          ifour       = _mm_set1_epi32(4);
1035     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
1036     real             *vftab;
1037     __m128d          dummy_mask,cutoff_mask;
1038     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1039     __m128d          one     = _mm_set1_pd(1.0);
1040     __m128d          two     = _mm_set1_pd(2.0);
1041     x                = xx[0];
1042     f                = ff[0];
1043
1044     nri              = nlist->nri;
1045     iinr             = nlist->iinr;
1046     jindex           = nlist->jindex;
1047     jjnr             = nlist->jjnr;
1048     shiftidx         = nlist->shift;
1049     gid              = nlist->gid;
1050     shiftvec         = fr->shift_vec[0];
1051     fshift           = fr->fshift[0];
1052     facel            = _mm_set1_pd(fr->epsfac);
1053     charge           = mdatoms->chargeA;
1054     nvdwtype         = fr->ntype;
1055     vdwparam         = fr->nbfp;
1056     vdwtype          = mdatoms->typeA;
1057
1058     vftab            = kernel_data->table_vdw->data;
1059     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
1060
1061     /* Setup water-specific parameters */
1062     inr              = nlist->iinr[0];
1063     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1064     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1065     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1066     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1067
1068     jq1              = _mm_set1_pd(charge[inr+1]);
1069     jq2              = _mm_set1_pd(charge[inr+2]);
1070     jq3              = _mm_set1_pd(charge[inr+3]);
1071     vdwjidx0A        = 2*vdwtype[inr+0];
1072     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1073     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1074     qq11             = _mm_mul_pd(iq1,jq1);
1075     qq12             = _mm_mul_pd(iq1,jq2);
1076     qq13             = _mm_mul_pd(iq1,jq3);
1077     qq21             = _mm_mul_pd(iq2,jq1);
1078     qq22             = _mm_mul_pd(iq2,jq2);
1079     qq23             = _mm_mul_pd(iq2,jq3);
1080     qq31             = _mm_mul_pd(iq3,jq1);
1081     qq32             = _mm_mul_pd(iq3,jq2);
1082     qq33             = _mm_mul_pd(iq3,jq3);
1083
1084     /* Avoid stupid compiler warnings */
1085     jnrA = jnrB = 0;
1086     j_coord_offsetA = 0;
1087     j_coord_offsetB = 0;
1088
1089     outeriter        = 0;
1090     inneriter        = 0;
1091
1092     /* Start outer loop over neighborlists */
1093     for(iidx=0; iidx<nri; iidx++)
1094     {
1095         /* Load shift vector for this list */
1096         i_shift_offset   = DIM*shiftidx[iidx];
1097
1098         /* Load limits for loop over neighbors */
1099         j_index_start    = jindex[iidx];
1100         j_index_end      = jindex[iidx+1];
1101
1102         /* Get outer coordinate index */
1103         inr              = iinr[iidx];
1104         i_coord_offset   = DIM*inr;
1105
1106         /* Load i particle coords and add shift vector */
1107         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1108                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1109
1110         fix0             = _mm_setzero_pd();
1111         fiy0             = _mm_setzero_pd();
1112         fiz0             = _mm_setzero_pd();
1113         fix1             = _mm_setzero_pd();
1114         fiy1             = _mm_setzero_pd();
1115         fiz1             = _mm_setzero_pd();
1116         fix2             = _mm_setzero_pd();
1117         fiy2             = _mm_setzero_pd();
1118         fiz2             = _mm_setzero_pd();
1119         fix3             = _mm_setzero_pd();
1120         fiy3             = _mm_setzero_pd();
1121         fiz3             = _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_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1135                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1136                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1137
1138             /* Calculate displacement vector */
1139             dx00             = _mm_sub_pd(ix0,jx0);
1140             dy00             = _mm_sub_pd(iy0,jy0);
1141             dz00             = _mm_sub_pd(iz0,jz0);
1142             dx11             = _mm_sub_pd(ix1,jx1);
1143             dy11             = _mm_sub_pd(iy1,jy1);
1144             dz11             = _mm_sub_pd(iz1,jz1);
1145             dx12             = _mm_sub_pd(ix1,jx2);
1146             dy12             = _mm_sub_pd(iy1,jy2);
1147             dz12             = _mm_sub_pd(iz1,jz2);
1148             dx13             = _mm_sub_pd(ix1,jx3);
1149             dy13             = _mm_sub_pd(iy1,jy3);
1150             dz13             = _mm_sub_pd(iz1,jz3);
1151             dx21             = _mm_sub_pd(ix2,jx1);
1152             dy21             = _mm_sub_pd(iy2,jy1);
1153             dz21             = _mm_sub_pd(iz2,jz1);
1154             dx22             = _mm_sub_pd(ix2,jx2);
1155             dy22             = _mm_sub_pd(iy2,jy2);
1156             dz22             = _mm_sub_pd(iz2,jz2);
1157             dx23             = _mm_sub_pd(ix2,jx3);
1158             dy23             = _mm_sub_pd(iy2,jy3);
1159             dz23             = _mm_sub_pd(iz2,jz3);
1160             dx31             = _mm_sub_pd(ix3,jx1);
1161             dy31             = _mm_sub_pd(iy3,jy1);
1162             dz31             = _mm_sub_pd(iz3,jz1);
1163             dx32             = _mm_sub_pd(ix3,jx2);
1164             dy32             = _mm_sub_pd(iy3,jy2);
1165             dz32             = _mm_sub_pd(iz3,jz2);
1166             dx33             = _mm_sub_pd(ix3,jx3);
1167             dy33             = _mm_sub_pd(iy3,jy3);
1168             dz33             = _mm_sub_pd(iz3,jz3);
1169
1170             /* Calculate squared distance and things based on it */
1171             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1172             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1173             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1174             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1175             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1176             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1177             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1178             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1179             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1180             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1181
1182             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1183             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1184             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1185             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1186             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1187             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1188             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1189             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1190             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1191             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1192
1193             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1194             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1195             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1196             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1197             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1198             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1199             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1200             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1201             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1202
1203             fjx0             = _mm_setzero_pd();
1204             fjy0             = _mm_setzero_pd();
1205             fjz0             = _mm_setzero_pd();
1206             fjx1             = _mm_setzero_pd();
1207             fjy1             = _mm_setzero_pd();
1208             fjz1             = _mm_setzero_pd();
1209             fjx2             = _mm_setzero_pd();
1210             fjy2             = _mm_setzero_pd();
1211             fjz2             = _mm_setzero_pd();
1212             fjx3             = _mm_setzero_pd();
1213             fjy3             = _mm_setzero_pd();
1214             fjz3             = _mm_setzero_pd();
1215
1216             /**************************
1217              * CALCULATE INTERACTIONS *
1218              **************************/
1219
1220             r00              = _mm_mul_pd(rsq00,rinv00);
1221
1222             /* Calculate table index by multiplying r with table scale and truncate to integer */
1223             rt               = _mm_mul_pd(r00,vftabscale);
1224             vfitab           = _mm_cvttpd_epi32(rt);
1225 #ifdef __XOP__
1226             vfeps            = _mm_frcz_pd(rt);
1227 #else
1228             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
1229 #endif
1230             twovfeps         = _mm_add_pd(vfeps,vfeps);
1231             vfitab           = _mm_slli_epi32(vfitab,3);
1232
1233             /* CUBIC SPLINE TABLE DISPERSION */
1234             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1235             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
1236             GMX_MM_TRANSPOSE2_PD(Y,F);
1237             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1238             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
1239             GMX_MM_TRANSPOSE2_PD(G,H);
1240             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1241             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1242             fvdw6            = _mm_mul_pd(c6_00,FF);
1243
1244             /* CUBIC SPLINE TABLE REPULSION */
1245             vfitab           = _mm_add_epi32(vfitab,ifour);
1246             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1247             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
1248             GMX_MM_TRANSPOSE2_PD(Y,F);
1249             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1250             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
1251             GMX_MM_TRANSPOSE2_PD(G,H);
1252             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1253             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1254             fvdw12           = _mm_mul_pd(c12_00,FF);
1255             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1256
1257             fscal            = fvdw;
1258
1259             /* Update vectorial force */
1260             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1261             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1262             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1263
1264             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1265             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1266             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1267
1268             /**************************
1269              * CALCULATE INTERACTIONS *
1270              **************************/
1271
1272             /* COULOMB ELECTROSTATICS */
1273             velec            = _mm_mul_pd(qq11,rinv11);
1274             felec            = _mm_mul_pd(velec,rinvsq11);
1275
1276             fscal            = felec;
1277
1278             /* Update vectorial force */
1279             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1280             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1281             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1282
1283             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1284             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1285             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1286
1287             /**************************
1288              * CALCULATE INTERACTIONS *
1289              **************************/
1290
1291             /* COULOMB ELECTROSTATICS */
1292             velec            = _mm_mul_pd(qq12,rinv12);
1293             felec            = _mm_mul_pd(velec,rinvsq12);
1294
1295             fscal            = felec;
1296
1297             /* Update vectorial force */
1298             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1299             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1300             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1301
1302             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1303             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1304             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1305
1306             /**************************
1307              * CALCULATE INTERACTIONS *
1308              **************************/
1309
1310             /* COULOMB ELECTROSTATICS */
1311             velec            = _mm_mul_pd(qq13,rinv13);
1312             felec            = _mm_mul_pd(velec,rinvsq13);
1313
1314             fscal            = felec;
1315
1316             /* Update vectorial force */
1317             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1318             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1319             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1320
1321             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1322             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1323             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1324
1325             /**************************
1326              * CALCULATE INTERACTIONS *
1327              **************************/
1328
1329             /* COULOMB ELECTROSTATICS */
1330             velec            = _mm_mul_pd(qq21,rinv21);
1331             felec            = _mm_mul_pd(velec,rinvsq21);
1332
1333             fscal            = felec;
1334
1335             /* Update vectorial force */
1336             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1337             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1338             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1339
1340             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1341             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1342             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1343
1344             /**************************
1345              * CALCULATE INTERACTIONS *
1346              **************************/
1347
1348             /* COULOMB ELECTROSTATICS */
1349             velec            = _mm_mul_pd(qq22,rinv22);
1350             felec            = _mm_mul_pd(velec,rinvsq22);
1351
1352             fscal            = felec;
1353
1354             /* Update vectorial force */
1355             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1356             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1357             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1358
1359             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1360             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1361             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1362
1363             /**************************
1364              * CALCULATE INTERACTIONS *
1365              **************************/
1366
1367             /* COULOMB ELECTROSTATICS */
1368             velec            = _mm_mul_pd(qq23,rinv23);
1369             felec            = _mm_mul_pd(velec,rinvsq23);
1370
1371             fscal            = felec;
1372
1373             /* Update vectorial force */
1374             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1375             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1376             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1377
1378             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1379             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1380             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1381
1382             /**************************
1383              * CALCULATE INTERACTIONS *
1384              **************************/
1385
1386             /* COULOMB ELECTROSTATICS */
1387             velec            = _mm_mul_pd(qq31,rinv31);
1388             felec            = _mm_mul_pd(velec,rinvsq31);
1389
1390             fscal            = felec;
1391
1392             /* Update vectorial force */
1393             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1394             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1395             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1396
1397             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1398             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1399             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1400
1401             /**************************
1402              * CALCULATE INTERACTIONS *
1403              **************************/
1404
1405             /* COULOMB ELECTROSTATICS */
1406             velec            = _mm_mul_pd(qq32,rinv32);
1407             felec            = _mm_mul_pd(velec,rinvsq32);
1408
1409             fscal            = felec;
1410
1411             /* Update vectorial force */
1412             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1413             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1414             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1415
1416             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1417             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1418             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1419
1420             /**************************
1421              * CALCULATE INTERACTIONS *
1422              **************************/
1423
1424             /* COULOMB ELECTROSTATICS */
1425             velec            = _mm_mul_pd(qq33,rinv33);
1426             felec            = _mm_mul_pd(velec,rinvsq33);
1427
1428             fscal            = felec;
1429
1430             /* Update vectorial force */
1431             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1432             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1433             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1434
1435             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1436             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1437             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1438
1439             gmx_mm_decrement_4rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1440
1441             /* Inner loop uses 324 flops */
1442         }
1443
1444         if(jidx<j_index_end)
1445         {
1446
1447             jnrA             = jjnr[jidx];
1448             j_coord_offsetA  = DIM*jnrA;
1449
1450             /* load j atom coordinates */
1451             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1452                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1453                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1454
1455             /* Calculate displacement vector */
1456             dx00             = _mm_sub_pd(ix0,jx0);
1457             dy00             = _mm_sub_pd(iy0,jy0);
1458             dz00             = _mm_sub_pd(iz0,jz0);
1459             dx11             = _mm_sub_pd(ix1,jx1);
1460             dy11             = _mm_sub_pd(iy1,jy1);
1461             dz11             = _mm_sub_pd(iz1,jz1);
1462             dx12             = _mm_sub_pd(ix1,jx2);
1463             dy12             = _mm_sub_pd(iy1,jy2);
1464             dz12             = _mm_sub_pd(iz1,jz2);
1465             dx13             = _mm_sub_pd(ix1,jx3);
1466             dy13             = _mm_sub_pd(iy1,jy3);
1467             dz13             = _mm_sub_pd(iz1,jz3);
1468             dx21             = _mm_sub_pd(ix2,jx1);
1469             dy21             = _mm_sub_pd(iy2,jy1);
1470             dz21             = _mm_sub_pd(iz2,jz1);
1471             dx22             = _mm_sub_pd(ix2,jx2);
1472             dy22             = _mm_sub_pd(iy2,jy2);
1473             dz22             = _mm_sub_pd(iz2,jz2);
1474             dx23             = _mm_sub_pd(ix2,jx3);
1475             dy23             = _mm_sub_pd(iy2,jy3);
1476             dz23             = _mm_sub_pd(iz2,jz3);
1477             dx31             = _mm_sub_pd(ix3,jx1);
1478             dy31             = _mm_sub_pd(iy3,jy1);
1479             dz31             = _mm_sub_pd(iz3,jz1);
1480             dx32             = _mm_sub_pd(ix3,jx2);
1481             dy32             = _mm_sub_pd(iy3,jy2);
1482             dz32             = _mm_sub_pd(iz3,jz2);
1483             dx33             = _mm_sub_pd(ix3,jx3);
1484             dy33             = _mm_sub_pd(iy3,jy3);
1485             dz33             = _mm_sub_pd(iz3,jz3);
1486
1487             /* Calculate squared distance and things based on it */
1488             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1489             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1490             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1491             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1492             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1493             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1494             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1495             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1496             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1497             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1498
1499             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1500             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1501             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1502             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1503             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1504             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1505             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1506             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1507             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1508             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1509
1510             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1511             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1512             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1513             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1514             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1515             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1516             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1517             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1518             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1519
1520             fjx0             = _mm_setzero_pd();
1521             fjy0             = _mm_setzero_pd();
1522             fjz0             = _mm_setzero_pd();
1523             fjx1             = _mm_setzero_pd();
1524             fjy1             = _mm_setzero_pd();
1525             fjz1             = _mm_setzero_pd();
1526             fjx2             = _mm_setzero_pd();
1527             fjy2             = _mm_setzero_pd();
1528             fjz2             = _mm_setzero_pd();
1529             fjx3             = _mm_setzero_pd();
1530             fjy3             = _mm_setzero_pd();
1531             fjz3             = _mm_setzero_pd();
1532
1533             /**************************
1534              * CALCULATE INTERACTIONS *
1535              **************************/
1536
1537             r00              = _mm_mul_pd(rsq00,rinv00);
1538
1539             /* Calculate table index by multiplying r with table scale and truncate to integer */
1540             rt               = _mm_mul_pd(r00,vftabscale);
1541             vfitab           = _mm_cvttpd_epi32(rt);
1542 #ifdef __XOP__
1543             vfeps            = _mm_frcz_pd(rt);
1544 #else
1545             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
1546 #endif
1547             twovfeps         = _mm_add_pd(vfeps,vfeps);
1548             vfitab           = _mm_slli_epi32(vfitab,3);
1549
1550             /* CUBIC SPLINE TABLE DISPERSION */
1551             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1552             F                = _mm_setzero_pd();
1553             GMX_MM_TRANSPOSE2_PD(Y,F);
1554             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1555             H                = _mm_setzero_pd();
1556             GMX_MM_TRANSPOSE2_PD(G,H);
1557             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1558             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1559             fvdw6            = _mm_mul_pd(c6_00,FF);
1560
1561             /* CUBIC SPLINE TABLE REPULSION */
1562             vfitab           = _mm_add_epi32(vfitab,ifour);
1563             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1564             F                = _mm_setzero_pd();
1565             GMX_MM_TRANSPOSE2_PD(Y,F);
1566             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1567             H                = _mm_setzero_pd();
1568             GMX_MM_TRANSPOSE2_PD(G,H);
1569             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1570             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1571             fvdw12           = _mm_mul_pd(c12_00,FF);
1572             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1573
1574             fscal            = fvdw;
1575
1576             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1577
1578             /* Update vectorial force */
1579             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1580             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1581             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1582
1583             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1584             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1585             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1586
1587             /**************************
1588              * CALCULATE INTERACTIONS *
1589              **************************/
1590
1591             /* COULOMB ELECTROSTATICS */
1592             velec            = _mm_mul_pd(qq11,rinv11);
1593             felec            = _mm_mul_pd(velec,rinvsq11);
1594
1595             fscal            = felec;
1596
1597             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1598
1599             /* Update vectorial force */
1600             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1601             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1602             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1603
1604             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1605             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1606             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1607
1608             /**************************
1609              * CALCULATE INTERACTIONS *
1610              **************************/
1611
1612             /* COULOMB ELECTROSTATICS */
1613             velec            = _mm_mul_pd(qq12,rinv12);
1614             felec            = _mm_mul_pd(velec,rinvsq12);
1615
1616             fscal            = felec;
1617
1618             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1619
1620             /* Update vectorial force */
1621             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1622             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1623             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1624
1625             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1626             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1627             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1628
1629             /**************************
1630              * CALCULATE INTERACTIONS *
1631              **************************/
1632
1633             /* COULOMB ELECTROSTATICS */
1634             velec            = _mm_mul_pd(qq13,rinv13);
1635             felec            = _mm_mul_pd(velec,rinvsq13);
1636
1637             fscal            = felec;
1638
1639             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1640
1641             /* Update vectorial force */
1642             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1643             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1644             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1645
1646             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1647             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1648             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1649
1650             /**************************
1651              * CALCULATE INTERACTIONS *
1652              **************************/
1653
1654             /* COULOMB ELECTROSTATICS */
1655             velec            = _mm_mul_pd(qq21,rinv21);
1656             felec            = _mm_mul_pd(velec,rinvsq21);
1657
1658             fscal            = felec;
1659
1660             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1661
1662             /* Update vectorial force */
1663             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1664             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1665             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1666
1667             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1668             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1669             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1670
1671             /**************************
1672              * CALCULATE INTERACTIONS *
1673              **************************/
1674
1675             /* COULOMB ELECTROSTATICS */
1676             velec            = _mm_mul_pd(qq22,rinv22);
1677             felec            = _mm_mul_pd(velec,rinvsq22);
1678
1679             fscal            = felec;
1680
1681             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1682
1683             /* Update vectorial force */
1684             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1685             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1686             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1687
1688             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1689             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1690             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1691
1692             /**************************
1693              * CALCULATE INTERACTIONS *
1694              **************************/
1695
1696             /* COULOMB ELECTROSTATICS */
1697             velec            = _mm_mul_pd(qq23,rinv23);
1698             felec            = _mm_mul_pd(velec,rinvsq23);
1699
1700             fscal            = felec;
1701
1702             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1703
1704             /* Update vectorial force */
1705             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1706             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1707             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1708
1709             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1710             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1711             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1712
1713             /**************************
1714              * CALCULATE INTERACTIONS *
1715              **************************/
1716
1717             /* COULOMB ELECTROSTATICS */
1718             velec            = _mm_mul_pd(qq31,rinv31);
1719             felec            = _mm_mul_pd(velec,rinvsq31);
1720
1721             fscal            = felec;
1722
1723             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1724
1725             /* Update vectorial force */
1726             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1727             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1728             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1729
1730             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1731             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1732             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1733
1734             /**************************
1735              * CALCULATE INTERACTIONS *
1736              **************************/
1737
1738             /* COULOMB ELECTROSTATICS */
1739             velec            = _mm_mul_pd(qq32,rinv32);
1740             felec            = _mm_mul_pd(velec,rinvsq32);
1741
1742             fscal            = felec;
1743
1744             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1745
1746             /* Update vectorial force */
1747             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1748             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1749             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1750
1751             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1752             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1753             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1754
1755             /**************************
1756              * CALCULATE INTERACTIONS *
1757              **************************/
1758
1759             /* COULOMB ELECTROSTATICS */
1760             velec            = _mm_mul_pd(qq33,rinv33);
1761             felec            = _mm_mul_pd(velec,rinvsq33);
1762
1763             fscal            = felec;
1764
1765             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1766
1767             /* Update vectorial force */
1768             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1769             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1770             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1771
1772             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1773             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1774             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1775
1776             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1777
1778             /* Inner loop uses 324 flops */
1779         }
1780
1781         /* End of innermost loop */
1782
1783         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1784                                               f+i_coord_offset,fshift+i_shift_offset);
1785
1786         /* Increment number of inner iterations */
1787         inneriter                  += j_index_end - j_index_start;
1788
1789         /* Outer loop uses 24 flops */
1790     }
1791
1792     /* Increment number of outer iterations */
1793     outeriter        += nri;
1794
1795     /* Update outer/inner flops */
1796
1797     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*324);
1798 }