src/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecCoul_VdwCSTab_GeomW4W4_avx_128_fma_double.c

   1 /*
   2  * Note: this file was generated by the Gromacs avx_128_fma_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_avx_128_fma_double.h"
  34 #include "kernelutil_x86_avx_128_fma_double.h"
  35
  36 /*
  37  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwCSTab_GeomW4W4_VF_avx_128_fma_double
  38  * Electrostatics interaction: Coulomb
  39  * VdW interaction:            CubicSplineTable
  40  * Geometry:                   Water4-Water4
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecCoul_VdwCSTab_GeomW4W4_VF_avx_128_fma_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              vdwioffset3;
  73     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  74     int              vdwjidx0A,vdwjidx0B;
  75     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  76     int              vdwjidx1A,vdwjidx1B;
  77     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  78     int              vdwjidx2A,vdwjidx2B;
  79     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  80     int              vdwjidx3A,vdwjidx3B;
  81     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  82     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  83     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  84     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  85     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
  86     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  87     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
  88     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
  89     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
  90     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
  91     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
  92     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
  93     real             *charge;
  94     int              nvdwtype;
  95     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
  96     int              *vdwtype;
  97     real             *vdwparam;
  98     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
  99     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 100     __m128i          vfitab;
 101     __m128i          ifour       = _mm_set1_epi32(4);
 102     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
 103     real             *vftab;
 104     __m128d          dummy_mask,cutoff_mask;
 105     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 106     __m128d          one     = _mm_set1_pd(1.0);
 107     __m128d          two     = _mm_set1_pd(2.0);
 108     x                = xx[0];
 109     f                = ff[0];
 110
 111     nri              = nlist->nri;
 112     iinr             = nlist->iinr;
 113     jindex           = nlist->jindex;
 114     jjnr             = nlist->jjnr;
 115     shiftidx         = nlist->shift;
 116     gid              = nlist->gid;
 117     shiftvec         = fr->shift_vec[0];
 118     fshift           = fr->fshift[0];
 119     facel            = _mm_set1_pd(fr->epsfac);
 120     charge           = mdatoms->chargeA;
 121     nvdwtype         = fr->ntype;
 122     vdwparam         = fr->nbfp;
 123     vdwtype          = mdatoms->typeA;
 124
 125     vftab            = kernel_data->table_vdw->data;
 126     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
 127
 128     /* Setup water-specific parameters */
 129     inr              = nlist->iinr[0];
 130     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 131     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 132     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 133     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 134
 135     jq1              = _mm_set1_pd(charge[inr+1]);
 136     jq2              = _mm_set1_pd(charge[inr+2]);
 137     jq3              = _mm_set1_pd(charge[inr+3]);
 138     vdwjidx0A        = 2*vdwtype[inr+0];
 139     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 140     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 141     qq11             = _mm_mul_pd(iq1,jq1);
 142     qq12             = _mm_mul_pd(iq1,jq2);
 143     qq13             = _mm_mul_pd(iq1,jq3);
 144     qq21             = _mm_mul_pd(iq2,jq1);
 145     qq22             = _mm_mul_pd(iq2,jq2);
 146     qq23             = _mm_mul_pd(iq2,jq3);
 147     qq31             = _mm_mul_pd(iq3,jq1);
 148     qq32             = _mm_mul_pd(iq3,jq2);
 149     qq33             = _mm_mul_pd(iq3,jq3);
 150
 151     /* Avoid stupid compiler warnings */
 152     jnrA = jnrB = 0;
 153     j_coord_offsetA = 0;
 154     j_coord_offsetB = 0;
 155
 156     outeriter        = 0;
 157     inneriter        = 0;
 158
 159     /* Start outer loop over neighborlists */
 160     for(iidx=0; iidx<nri; iidx++)
 161     {
 162         /* Load shift vector for this list */
 163         i_shift_offset   = DIM*shiftidx[iidx];
 164
 165         /* Load limits for loop over neighbors */
 166         j_index_start    = jindex[iidx];
 167         j_index_end      = jindex[iidx+1];
 168
 169         /* Get outer coordinate index */
 170         inr              = iinr[iidx];
 171         i_coord_offset   = DIM*inr;
 172
 173         /* Load i particle coords and add shift vector */
 174         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 175                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 176
 177         fix0             = _mm_setzero_pd();
 178         fiy0             = _mm_setzero_pd();
 179         fiz0             = _mm_setzero_pd();
 180         fix1             = _mm_setzero_pd();
 181         fiy1             = _mm_setzero_pd();
 182         fiz1             = _mm_setzero_pd();
 183         fix2             = _mm_setzero_pd();
 184         fiy2             = _mm_setzero_pd();
 185         fiz2             = _mm_setzero_pd();
 186         fix3             = _mm_setzero_pd();
 187         fiy3             = _mm_setzero_pd();
 188         fiz3             = _mm_setzero_pd();
 189
 190         /* Reset potential sums */
 191         velecsum         = _mm_setzero_pd();
 192         vvdwsum          = _mm_setzero_pd();
 193
 194         /* Start inner kernel loop */
 195         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 196         {
 197
 198             /* Get j neighbor index, and coordinate index */
 199             jnrA             = jjnr[jidx];
 200             jnrB             = jjnr[jidx+1];
 201             j_coord_offsetA  = DIM*jnrA;
 202             j_coord_offsetB  = DIM*jnrB;
 203
 204             /* load j atom coordinates */
 205             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 206                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 207                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 208
 209             /* Calculate displacement vector */
 210             dx00             = _mm_sub_pd(ix0,jx0);
 211             dy00             = _mm_sub_pd(iy0,jy0);
 212             dz00             = _mm_sub_pd(iz0,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             dx13             = _mm_sub_pd(ix1,jx3);
 220             dy13             = _mm_sub_pd(iy1,jy3);
 221             dz13             = _mm_sub_pd(iz1,jz3);
 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             dx23             = _mm_sub_pd(ix2,jx3);
 229             dy23             = _mm_sub_pd(iy2,jy3);
 230             dz23             = _mm_sub_pd(iz2,jz3);
 231             dx31             = _mm_sub_pd(ix3,jx1);
 232             dy31             = _mm_sub_pd(iy3,jy1);
 233             dz31             = _mm_sub_pd(iz3,jz1);
 234             dx32             = _mm_sub_pd(ix3,jx2);
 235             dy32             = _mm_sub_pd(iy3,jy2);
 236             dz32             = _mm_sub_pd(iz3,jz2);
 237             dx33             = _mm_sub_pd(ix3,jx3);
 238             dy33             = _mm_sub_pd(iy3,jy3);
 239             dz33             = _mm_sub_pd(iz3,jz3);
 240
 241             /* Calculate squared distance and things based on it */
 242             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 243             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 244             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 245             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 246             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 247             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 248             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 249             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 250             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 251             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 252
 253             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 254             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 255             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 256             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 257             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 258             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 259             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 260             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 261             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 262             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 263
 264             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 265             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 266             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 267             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 268             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 269             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 270             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 271             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 272             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 273
 274             fjx0             = _mm_setzero_pd();
 275             fjy0             = _mm_setzero_pd();
 276             fjz0             = _mm_setzero_pd();
 277             fjx1             = _mm_setzero_pd();
 278             fjy1             = _mm_setzero_pd();
 279             fjz1             = _mm_setzero_pd();
 280             fjx2             = _mm_setzero_pd();
 281             fjy2             = _mm_setzero_pd();
 282             fjz2             = _mm_setzero_pd();
 283             fjx3             = _mm_setzero_pd();
 284             fjy3             = _mm_setzero_pd();
 285             fjz3             = _mm_setzero_pd();
 286
 287             /**************************
 288              * CALCULATE INTERACTIONS *
 289              **************************/
 290
 291             r00              = _mm_mul_pd(rsq00,rinv00);
 292
 293             /* Calculate table index by multiplying r with table scale and truncate to integer */
 294             rt               = _mm_mul_pd(r00,vftabscale);
 295             vfitab           = _mm_cvttpd_epi32(rt);
 296 #ifdef __XOP__
 297             vfeps            = _mm_frcz_pd(rt);
 298 #else
 299             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 300 #endif
 301             twovfeps         = _mm_add_pd(vfeps,vfeps);
 302             vfitab           = _mm_slli_epi32(vfitab,3);
 303
 304             /* CUBIC SPLINE TABLE DISPERSION */
 305             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 306             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 307             GMX_MM_TRANSPOSE2_PD(Y,F);
 308             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 309             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
 310             GMX_MM_TRANSPOSE2_PD(G,H);
 311             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 312             VV               = _mm_macc_pd(vfeps,Fp,Y);
 313             vvdw6            = _mm_mul_pd(c6_00,VV);
 314             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 315             fvdw6            = _mm_mul_pd(c6_00,FF);
 316
 317             /* CUBIC SPLINE TABLE REPULSION */
 318             vfitab           = _mm_add_epi32(vfitab,ifour);
 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             vvdw12           = _mm_mul_pd(c12_00,VV);
 328             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 329             fvdw12           = _mm_mul_pd(c12_00,FF);
 330             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 331             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 332
 333             /* Update potential sum for this i atom from the interaction with this j atom. */
 334             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 335
 336             fscal            = fvdw;
 337
 338             /* Update vectorial force */
 339             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 340             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 341             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 342
 343             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 344             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 345             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 346
 347             /**************************
 348              * CALCULATE INTERACTIONS *
 349              **************************/
 350
 351             /* COULOMB ELECTROSTATICS */
 352             velec            = _mm_mul_pd(qq11,rinv11);
 353             felec            = _mm_mul_pd(velec,rinvsq11);
 354
 355             /* Update potential sum for this i atom from the interaction with this j atom. */
 356             velecsum         = _mm_add_pd(velecsum,velec);
 357
 358             fscal            = felec;
 359
 360             /* Update vectorial force */
 361             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 362             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 363             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 364
 365             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 366             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 367             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 368
 369             /**************************
 370              * CALCULATE INTERACTIONS *
 371              **************************/
 372
 373             /* COULOMB ELECTROSTATICS */
 374             velec            = _mm_mul_pd(qq12,rinv12);
 375             felec            = _mm_mul_pd(velec,rinvsq12);
 376
 377             /* Update potential sum for this i atom from the interaction with this j atom. */
 378             velecsum         = _mm_add_pd(velecsum,velec);
 379
 380             fscal            = felec;
 381
 382             /* Update vectorial force */
 383             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 384             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 385             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 386
 387             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 388             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 389             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 390
 391             /**************************
 392              * CALCULATE INTERACTIONS *
 393              **************************/
 394
 395             /* COULOMB ELECTROSTATICS */
 396             velec            = _mm_mul_pd(qq13,rinv13);
 397             felec            = _mm_mul_pd(velec,rinvsq13);
 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             /* Update vectorial force */
 405             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 406             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 407             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 408
 409             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 410             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 411             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 412
 413             /**************************
 414              * CALCULATE INTERACTIONS *
 415              **************************/
 416
 417             /* COULOMB ELECTROSTATICS */
 418             velec            = _mm_mul_pd(qq21,rinv21);
 419             felec            = _mm_mul_pd(velec,rinvsq21);
 420
 421             /* Update potential sum for this i atom from the interaction with this j atom. */
 422             velecsum         = _mm_add_pd(velecsum,velec);
 423
 424             fscal            = felec;
 425
 426             /* Update vectorial force */
 427             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 428             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 429             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 430
 431             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 432             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 433             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 434
 435             /**************************
 436              * CALCULATE INTERACTIONS *
 437              **************************/
 438
 439             /* COULOMB ELECTROSTATICS */
 440             velec            = _mm_mul_pd(qq22,rinv22);
 441             felec            = _mm_mul_pd(velec,rinvsq22);
 442
 443             /* Update potential sum for this i atom from the interaction with this j atom. */
 444             velecsum         = _mm_add_pd(velecsum,velec);
 445
 446             fscal            = felec;
 447
 448             /* Update vectorial force */
 449             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 450             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 451             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 452
 453             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 454             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 455             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 456
 457             /**************************
 458              * CALCULATE INTERACTIONS *
 459              **************************/
 460
 461             /* COULOMB ELECTROSTATICS */
 462             velec            = _mm_mul_pd(qq23,rinv23);
 463             felec            = _mm_mul_pd(velec,rinvsq23);
 464
 465             /* Update potential sum for this i atom from the interaction with this j atom. */
 466             velecsum         = _mm_add_pd(velecsum,velec);
 467
 468             fscal            = felec;
 469
 470             /* Update vectorial force */
 471             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 472             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 473             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 474
 475             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 476             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 477             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 478
 479             /**************************
 480              * CALCULATE INTERACTIONS *
 481              **************************/
 482
 483             /* COULOMB ELECTROSTATICS */
 484             velec            = _mm_mul_pd(qq31,rinv31);
 485             felec            = _mm_mul_pd(velec,rinvsq31);
 486
 487             /* Update potential sum for this i atom from the interaction with this j atom. */
 488             velecsum         = _mm_add_pd(velecsum,velec);
 489
 490             fscal            = felec;
 491
 492             /* Update vectorial force */
 493             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 494             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 495             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 496
 497             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 498             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 499             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 500
 501             /**************************
 502              * CALCULATE INTERACTIONS *
 503              **************************/
 504
 505             /* COULOMB ELECTROSTATICS */
 506             velec            = _mm_mul_pd(qq32,rinv32);
 507             felec            = _mm_mul_pd(velec,rinvsq32);
 508
 509             /* Update potential sum for this i atom from the interaction with this j atom. */
 510             velecsum         = _mm_add_pd(velecsum,velec);
 511
 512             fscal            = felec;
 513
 514             /* Update vectorial force */
 515             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 516             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 517             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 518
 519             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 520             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 521             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 522
 523             /**************************
 524              * CALCULATE INTERACTIONS *
 525              **************************/
 526
 527             /* COULOMB ELECTROSTATICS */
 528             velec            = _mm_mul_pd(qq33,rinv33);
 529             felec            = _mm_mul_pd(velec,rinvsq33);
 530
 531             /* Update potential sum for this i atom from the interaction with this j atom. */
 532             velecsum         = _mm_add_pd(velecsum,velec);
 533
 534             fscal            = felec;
 535
 536             /* Update vectorial force */
 537             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 538             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 539             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 540
 541             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 542             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 543             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 544
 545             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);
 546
 547             /* Inner loop uses 341 flops */
 548         }
 549
 550         if(jidx<j_index_end)
 551         {
 552
 553             jnrA             = jjnr[jidx];
 554             j_coord_offsetA  = DIM*jnrA;
 555
 556             /* load j atom coordinates */
 557             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 558                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 559                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 560
 561             /* Calculate displacement vector */
 562             dx00             = _mm_sub_pd(ix0,jx0);
 563             dy00             = _mm_sub_pd(iy0,jy0);
 564             dz00             = _mm_sub_pd(iz0,jz0);
 565             dx11             = _mm_sub_pd(ix1,jx1);
 566             dy11             = _mm_sub_pd(iy1,jy1);
 567             dz11             = _mm_sub_pd(iz1,jz1);
 568             dx12             = _mm_sub_pd(ix1,jx2);
 569             dy12             = _mm_sub_pd(iy1,jy2);
 570             dz12             = _mm_sub_pd(iz1,jz2);
 571             dx13             = _mm_sub_pd(ix1,jx3);
 572             dy13             = _mm_sub_pd(iy1,jy3);
 573             dz13             = _mm_sub_pd(iz1,jz3);
 574             dx21             = _mm_sub_pd(ix2,jx1);
 575             dy21             = _mm_sub_pd(iy2,jy1);
 576             dz21             = _mm_sub_pd(iz2,jz1);
 577             dx22             = _mm_sub_pd(ix2,jx2);
 578             dy22             = _mm_sub_pd(iy2,jy2);
 579             dz22             = _mm_sub_pd(iz2,jz2);
 580             dx23             = _mm_sub_pd(ix2,jx3);
 581             dy23             = _mm_sub_pd(iy2,jy3);
 582             dz23             = _mm_sub_pd(iz2,jz3);
 583             dx31             = _mm_sub_pd(ix3,jx1);
 584             dy31             = _mm_sub_pd(iy3,jy1);
 585             dz31             = _mm_sub_pd(iz3,jz1);
 586             dx32             = _mm_sub_pd(ix3,jx2);
 587             dy32             = _mm_sub_pd(iy3,jy2);
 588             dz32             = _mm_sub_pd(iz3,jz2);
 589             dx33             = _mm_sub_pd(ix3,jx3);
 590             dy33             = _mm_sub_pd(iy3,jy3);
 591             dz33             = _mm_sub_pd(iz3,jz3);
 592
 593             /* Calculate squared distance and things based on it */
 594             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 595             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 596             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 597             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 598             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 599             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 600             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 601             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 602             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 603             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 604
 605             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 606             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 607             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 608             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 609             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 610             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 611             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 612             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 613             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 614             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 615
 616             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 617             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 618             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 619             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 620             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 621             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 622             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 623             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 624             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 625
 626             fjx0             = _mm_setzero_pd();
 627             fjy0             = _mm_setzero_pd();
 628             fjz0             = _mm_setzero_pd();
 629             fjx1             = _mm_setzero_pd();
 630             fjy1             = _mm_setzero_pd();
 631             fjz1             = _mm_setzero_pd();
 632             fjx2             = _mm_setzero_pd();
 633             fjy2             = _mm_setzero_pd();
 634             fjz2             = _mm_setzero_pd();
 635             fjx3             = _mm_setzero_pd();
 636             fjy3             = _mm_setzero_pd();
 637             fjz3             = _mm_setzero_pd();
 638
 639             /**************************
 640              * CALCULATE INTERACTIONS *
 641              **************************/
 642
 643             r00              = _mm_mul_pd(rsq00,rinv00);
 644
 645             /* Calculate table index by multiplying r with table scale and truncate to integer */
 646             rt               = _mm_mul_pd(r00,vftabscale);
 647             vfitab           = _mm_cvttpd_epi32(rt);
 648 #ifdef __XOP__
 649             vfeps            = _mm_frcz_pd(rt);
 650 #else
 651             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 652 #endif
 653             twovfeps         = _mm_add_pd(vfeps,vfeps);
 654             vfitab           = _mm_slli_epi32(vfitab,3);
 655
 656             /* CUBIC SPLINE TABLE DISPERSION */
 657             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 658             F                = _mm_setzero_pd();
 659             GMX_MM_TRANSPOSE2_PD(Y,F);
 660             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 661             H                = _mm_setzero_pd();
 662             GMX_MM_TRANSPOSE2_PD(G,H);
 663             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 664             VV               = _mm_macc_pd(vfeps,Fp,Y);
 665             vvdw6            = _mm_mul_pd(c6_00,VV);
 666             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 667             fvdw6            = _mm_mul_pd(c6_00,FF);
 668
 669             /* CUBIC SPLINE TABLE REPULSION */
 670             vfitab           = _mm_add_epi32(vfitab,ifour);
 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             vvdw12           = _mm_mul_pd(c12_00,VV);
 680             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 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             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 687             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 688
 689             fscal            = fvdw;
 690
 691             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 692
 693             /* Update vectorial force */
 694             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 695             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 696             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 697
 698             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 699             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 700             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 701
 702             /**************************
 703              * CALCULATE INTERACTIONS *
 704              **************************/
 705
 706             /* COULOMB ELECTROSTATICS */
 707             velec            = _mm_mul_pd(qq11,rinv11);
 708             felec            = _mm_mul_pd(velec,rinvsq11);
 709
 710             /* Update potential sum for this i atom from the interaction with this j atom. */
 711             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 712             velecsum         = _mm_add_pd(velecsum,velec);
 713
 714             fscal            = felec;
 715
 716             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 717
 718             /* Update vectorial force */
 719             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 720             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 721             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 722
 723             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 724             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 725             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 726
 727             /**************************
 728              * CALCULATE INTERACTIONS *
 729              **************************/
 730
 731             /* COULOMB ELECTROSTATICS */
 732             velec            = _mm_mul_pd(qq12,rinv12);
 733             felec            = _mm_mul_pd(velec,rinvsq12);
 734
 735             /* Update potential sum for this i atom from the interaction with this j atom. */
 736             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 737             velecsum         = _mm_add_pd(velecsum,velec);
 738
 739             fscal            = felec;
 740
 741             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 742
 743             /* Update vectorial force */
 744             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 745             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 746             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 747
 748             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 749             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 750             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 751
 752             /**************************
 753              * CALCULATE INTERACTIONS *
 754              **************************/
 755
 756             /* COULOMB ELECTROSTATICS */
 757             velec            = _mm_mul_pd(qq13,rinv13);
 758             felec            = _mm_mul_pd(velec,rinvsq13);
 759
 760             /* Update potential sum for this i atom from the interaction with this j atom. */
 761             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 762             velecsum         = _mm_add_pd(velecsum,velec);
 763
 764             fscal            = felec;
 765
 766             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 767
 768             /* Update vectorial force */
 769             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 770             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 771             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 772
 773             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 774             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 775             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 776
 777             /**************************
 778              * CALCULATE INTERACTIONS *
 779              **************************/
 780
 781             /* COULOMB ELECTROSTATICS */
 782             velec            = _mm_mul_pd(qq21,rinv21);
 783             felec            = _mm_mul_pd(velec,rinvsq21);
 784
 785             /* Update potential sum for this i atom from the interaction with this j atom. */
 786             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 787             velecsum         = _mm_add_pd(velecsum,velec);
 788
 789             fscal            = felec;
 790
 791             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 792
 793             /* Update vectorial force */
 794             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 795             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 796             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 797
 798             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 799             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 800             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 801
 802             /**************************
 803              * CALCULATE INTERACTIONS *
 804              **************************/
 805
 806             /* COULOMB ELECTROSTATICS */
 807             velec            = _mm_mul_pd(qq22,rinv22);
 808             felec            = _mm_mul_pd(velec,rinvsq22);
 809
 810             /* Update potential sum for this i atom from the interaction with this j atom. */
 811             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 812             velecsum         = _mm_add_pd(velecsum,velec);
 813
 814             fscal            = felec;
 815
 816             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 817
 818             /* Update vectorial force */
 819             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 820             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 821             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 822
 823             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 824             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 825             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 826
 827             /**************************
 828              * CALCULATE INTERACTIONS *
 829              **************************/
 830
 831             /* COULOMB ELECTROSTATICS */
 832             velec            = _mm_mul_pd(qq23,rinv23);
 833             felec            = _mm_mul_pd(velec,rinvsq23);
 834
 835             /* Update potential sum for this i atom from the interaction with this j atom. */
 836             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 837             velecsum         = _mm_add_pd(velecsum,velec);
 838
 839             fscal            = felec;
 840
 841             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 842
 843             /* Update vectorial force */
 844             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 845             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 846             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 847
 848             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 849             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 850             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 851
 852             /**************************
 853              * CALCULATE INTERACTIONS *
 854              **************************/
 855
 856             /* COULOMB ELECTROSTATICS */
 857             velec            = _mm_mul_pd(qq31,rinv31);
 858             felec            = _mm_mul_pd(velec,rinvsq31);
 859
 860             /* Update potential sum for this i atom from the interaction with this j atom. */
 861             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 862             velecsum         = _mm_add_pd(velecsum,velec);
 863
 864             fscal            = felec;
 865
 866             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 867
 868             /* Update vectorial force */
 869             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 870             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 871             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 872
 873             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 874             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 875             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 876
 877             /**************************
 878              * CALCULATE INTERACTIONS *
 879              **************************/
 880
 881             /* COULOMB ELECTROSTATICS */
 882             velec            = _mm_mul_pd(qq32,rinv32);
 883             felec            = _mm_mul_pd(velec,rinvsq32);
 884
 885             /* Update potential sum for this i atom from the interaction with this j atom. */
 886             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 887             velecsum         = _mm_add_pd(velecsum,velec);
 888
 889             fscal            = felec;
 890
 891             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 892
 893             /* Update vectorial force */
 894             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 895             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 896             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 897
 898             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 899             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 900             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 901
 902             /**************************
 903              * CALCULATE INTERACTIONS *
 904              **************************/
 905
 906             /* COULOMB ELECTROSTATICS */
 907             velec            = _mm_mul_pd(qq33,rinv33);
 908             felec            = _mm_mul_pd(velec,rinvsq33);
 909
 910             /* Update potential sum for this i atom from the interaction with this j atom. */
 911             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 912             velecsum         = _mm_add_pd(velecsum,velec);
 913
 914             fscal            = felec;
 915
 916             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 917
 918             /* Update vectorial force */
 919             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 920             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 921             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 922
 923             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 924             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 925             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 926
 927             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 928
 929             /* Inner loop uses 341 flops */
 930         }
 931
 932         /* End of innermost loop */
 933
 934         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 935                                               f+i_coord_offset,fshift+i_shift_offset);
 936
 937         ggid                        = gid[iidx];
 938         /* Update potential energies */
 939         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 940         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 941
 942         /* Increment number of inner iterations */
 943         inneriter                  += j_index_end - j_index_start;
 944
 945         /* Outer loop uses 26 flops */
 946     }
 947
 948     /* Increment number of outer iterations */
 949     outeriter        += nri;
 950
 951     /* Update outer/inner flops */
 952
 953     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*341);
 954 }
 955 /*
 956  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwCSTab_GeomW4W4_F_avx_128_fma_double
 957  * Electrostatics interaction: Coulomb
 958  * VdW interaction:            CubicSplineTable
 959  * Geometry:                   Water4-Water4
 960  * Calculate force/pot:        Force
 961  */
 962 void
 963 nb_kernel_ElecCoul_VdwCSTab_GeomW4W4_F_avx_128_fma_double
 964                     (t_nblist * gmx_restrict                nlist,
 965                      rvec * gmx_restrict                    xx,
 966                      rvec * gmx_restrict                    ff,
 967                      t_forcerec * gmx_restrict              fr,
 968                      t_mdatoms * gmx_restrict               mdatoms,
 969                      nb_kernel_data_t * gmx_restrict        kernel_data,
 970                      t_nrnb * gmx_restrict                  nrnb)
 971 {
 972     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 973      * just 0 for non-waters.
 974      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 975      * jnr indices corresponding to data put in the four positions in the SIMD register.
 976      */
 977     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 978     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 979     int              jnrA,jnrB;
 980     int              j_coord_offsetA,j_coord_offsetB;
 981     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 982     real             rcutoff_scalar;
 983     real             *shiftvec,*fshift,*x,*f;
 984     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 985     int              vdwioffset0;
 986     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 987     int              vdwioffset1;
 988     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 989     int              vdwioffset2;
 990     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 991     int              vdwioffset3;
 992     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 993     int              vdwjidx0A,vdwjidx0B;
 994     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 995     int              vdwjidx1A,vdwjidx1B;
 996     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 997     int              vdwjidx2A,vdwjidx2B;
 998     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 999     int              vdwjidx3A,vdwjidx3B;
1000     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1001     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1002     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1003     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1004     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1005     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1006     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1007     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1008     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1009     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1010     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1011     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1012     real             *charge;
1013     int              nvdwtype;
1014     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1015     int              *vdwtype;
1016     real             *vdwparam;
1017     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1018     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1019     __m128i          vfitab;
1020     __m128i          ifour       = _mm_set1_epi32(4);
1021     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
1022     real             *vftab;
1023     __m128d          dummy_mask,cutoff_mask;
1024     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1025     __m128d          one     = _mm_set1_pd(1.0);
1026     __m128d          two     = _mm_set1_pd(2.0);
1027     x                = xx[0];
1028     f                = ff[0];
1029
1030     nri              = nlist->nri;
1031     iinr             = nlist->iinr;
1032     jindex           = nlist->jindex;
1033     jjnr             = nlist->jjnr;
1034     shiftidx         = nlist->shift;
1035     gid              = nlist->gid;
1036     shiftvec         = fr->shift_vec[0];
1037     fshift           = fr->fshift[0];
1038     facel            = _mm_set1_pd(fr->epsfac);
1039     charge           = mdatoms->chargeA;
1040     nvdwtype         = fr->ntype;
1041     vdwparam         = fr->nbfp;
1042     vdwtype          = mdatoms->typeA;
1043
1044     vftab            = kernel_data->table_vdw->data;
1045     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
1046
1047     /* Setup water-specific parameters */
1048     inr              = nlist->iinr[0];
1049     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1050     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1051     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1052     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1053
1054     jq1              = _mm_set1_pd(charge[inr+1]);
1055     jq2              = _mm_set1_pd(charge[inr+2]);
1056     jq3              = _mm_set1_pd(charge[inr+3]);
1057     vdwjidx0A        = 2*vdwtype[inr+0];
1058     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1059     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1060     qq11             = _mm_mul_pd(iq1,jq1);
1061     qq12             = _mm_mul_pd(iq1,jq2);
1062     qq13             = _mm_mul_pd(iq1,jq3);
1063     qq21             = _mm_mul_pd(iq2,jq1);
1064     qq22             = _mm_mul_pd(iq2,jq2);
1065     qq23             = _mm_mul_pd(iq2,jq3);
1066     qq31             = _mm_mul_pd(iq3,jq1);
1067     qq32             = _mm_mul_pd(iq3,jq2);
1068     qq33             = _mm_mul_pd(iq3,jq3);
1069
1070     /* Avoid stupid compiler warnings */
1071     jnrA = jnrB = 0;
1072     j_coord_offsetA = 0;
1073     j_coord_offsetB = 0;
1074
1075     outeriter        = 0;
1076     inneriter        = 0;
1077
1078     /* Start outer loop over neighborlists */
1079     for(iidx=0; iidx<nri; iidx++)
1080     {
1081         /* Load shift vector for this list */
1082         i_shift_offset   = DIM*shiftidx[iidx];
1083
1084         /* Load limits for loop over neighbors */
1085         j_index_start    = jindex[iidx];
1086         j_index_end      = jindex[iidx+1];
1087
1088         /* Get outer coordinate index */
1089         inr              = iinr[iidx];
1090         i_coord_offset   = DIM*inr;
1091
1092         /* Load i particle coords and add shift vector */
1093         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1094                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1095
1096         fix0             = _mm_setzero_pd();
1097         fiy0             = _mm_setzero_pd();
1098         fiz0             = _mm_setzero_pd();
1099         fix1             = _mm_setzero_pd();
1100         fiy1             = _mm_setzero_pd();
1101         fiz1             = _mm_setzero_pd();
1102         fix2             = _mm_setzero_pd();
1103         fiy2             = _mm_setzero_pd();
1104         fiz2             = _mm_setzero_pd();
1105         fix3             = _mm_setzero_pd();
1106         fiy3             = _mm_setzero_pd();
1107         fiz3             = _mm_setzero_pd();
1108
1109         /* Start inner kernel loop */
1110         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1111         {
1112
1113             /* Get j neighbor index, and coordinate index */
1114             jnrA             = jjnr[jidx];
1115             jnrB             = jjnr[jidx+1];
1116             j_coord_offsetA  = DIM*jnrA;
1117             j_coord_offsetB  = DIM*jnrB;
1118
1119             /* load j atom coordinates */
1120             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1121                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1122                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1123
1124             /* Calculate displacement vector */
1125             dx00             = _mm_sub_pd(ix0,jx0);
1126             dy00             = _mm_sub_pd(iy0,jy0);
1127             dz00             = _mm_sub_pd(iz0,jz0);
1128             dx11             = _mm_sub_pd(ix1,jx1);
1129             dy11             = _mm_sub_pd(iy1,jy1);
1130             dz11             = _mm_sub_pd(iz1,jz1);
1131             dx12             = _mm_sub_pd(ix1,jx2);
1132             dy12             = _mm_sub_pd(iy1,jy2);
1133             dz12             = _mm_sub_pd(iz1,jz2);
1134             dx13             = _mm_sub_pd(ix1,jx3);
1135             dy13             = _mm_sub_pd(iy1,jy3);
1136             dz13             = _mm_sub_pd(iz1,jz3);
1137             dx21             = _mm_sub_pd(ix2,jx1);
1138             dy21             = _mm_sub_pd(iy2,jy1);
1139             dz21             = _mm_sub_pd(iz2,jz1);
1140             dx22             = _mm_sub_pd(ix2,jx2);
1141             dy22             = _mm_sub_pd(iy2,jy2);
1142             dz22             = _mm_sub_pd(iz2,jz2);
1143             dx23             = _mm_sub_pd(ix2,jx3);
1144             dy23             = _mm_sub_pd(iy2,jy3);
1145             dz23             = _mm_sub_pd(iz2,jz3);
1146             dx31             = _mm_sub_pd(ix3,jx1);
1147             dy31             = _mm_sub_pd(iy3,jy1);
1148             dz31             = _mm_sub_pd(iz3,jz1);
1149             dx32             = _mm_sub_pd(ix3,jx2);
1150             dy32             = _mm_sub_pd(iy3,jy2);
1151             dz32             = _mm_sub_pd(iz3,jz2);
1152             dx33             = _mm_sub_pd(ix3,jx3);
1153             dy33             = _mm_sub_pd(iy3,jy3);
1154             dz33             = _mm_sub_pd(iz3,jz3);
1155
1156             /* Calculate squared distance and things based on it */
1157             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1158             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1159             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1160             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1161             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1162             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1163             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1164             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1165             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1166             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1167
1168             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1169             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1170             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1171             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1172             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1173             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1174             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1175             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1176             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1177             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1178
1179             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1180             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1181             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1182             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1183             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1184             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1185             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1186             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1187             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1188
1189             fjx0             = _mm_setzero_pd();
1190             fjy0             = _mm_setzero_pd();
1191             fjz0             = _mm_setzero_pd();
1192             fjx1             = _mm_setzero_pd();
1193             fjy1             = _mm_setzero_pd();
1194             fjz1             = _mm_setzero_pd();
1195             fjx2             = _mm_setzero_pd();
1196             fjy2             = _mm_setzero_pd();
1197             fjz2             = _mm_setzero_pd();
1198             fjx3             = _mm_setzero_pd();
1199             fjy3             = _mm_setzero_pd();
1200             fjz3             = _mm_setzero_pd();
1201
1202             /**************************
1203              * CALCULATE INTERACTIONS *
1204              **************************/
1205
1206             r00              = _mm_mul_pd(rsq00,rinv00);
1207
1208             /* Calculate table index by multiplying r with table scale and truncate to integer */
1209             rt               = _mm_mul_pd(r00,vftabscale);
1210             vfitab           = _mm_cvttpd_epi32(rt);
1211 #ifdef __XOP__
1212             vfeps            = _mm_frcz_pd(rt);
1213 #else
1214             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
1215 #endif
1216             twovfeps         = _mm_add_pd(vfeps,vfeps);
1217             vfitab           = _mm_slli_epi32(vfitab,3);
1218
1219             /* CUBIC SPLINE TABLE DISPERSION */
1220             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1221             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
1222             GMX_MM_TRANSPOSE2_PD(Y,F);
1223             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1224             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
1225             GMX_MM_TRANSPOSE2_PD(G,H);
1226             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1227             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1228             fvdw6            = _mm_mul_pd(c6_00,FF);
1229
1230             /* CUBIC SPLINE TABLE REPULSION */
1231             vfitab           = _mm_add_epi32(vfitab,ifour);
1232             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1233             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
1234             GMX_MM_TRANSPOSE2_PD(Y,F);
1235             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1236             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
1237             GMX_MM_TRANSPOSE2_PD(G,H);
1238             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1239             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1240             fvdw12           = _mm_mul_pd(c12_00,FF);
1241             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1242
1243             fscal            = fvdw;
1244
1245             /* Update vectorial force */
1246             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1247             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1248             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1249
1250             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1251             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1252             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1253
1254             /**************************
1255              * CALCULATE INTERACTIONS *
1256              **************************/
1257
1258             /* COULOMB ELECTROSTATICS */
1259             velec            = _mm_mul_pd(qq11,rinv11);
1260             felec            = _mm_mul_pd(velec,rinvsq11);
1261
1262             fscal            = felec;
1263
1264             /* Update vectorial force */
1265             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1266             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1267             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1268
1269             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1270             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1271             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1272
1273             /**************************
1274              * CALCULATE INTERACTIONS *
1275              **************************/
1276
1277             /* COULOMB ELECTROSTATICS */
1278             velec            = _mm_mul_pd(qq12,rinv12);
1279             felec            = _mm_mul_pd(velec,rinvsq12);
1280
1281             fscal            = felec;
1282
1283             /* Update vectorial force */
1284             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1285             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1286             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1287
1288             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1289             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1290             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1291
1292             /**************************
1293              * CALCULATE INTERACTIONS *
1294              **************************/
1295
1296             /* COULOMB ELECTROSTATICS */
1297             velec            = _mm_mul_pd(qq13,rinv13);
1298             felec            = _mm_mul_pd(velec,rinvsq13);
1299
1300             fscal            = felec;
1301
1302             /* Update vectorial force */
1303             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1304             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1305             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1306
1307             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1308             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1309             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1310
1311             /**************************
1312              * CALCULATE INTERACTIONS *
1313              **************************/
1314
1315             /* COULOMB ELECTROSTATICS */
1316             velec            = _mm_mul_pd(qq21,rinv21);
1317             felec            = _mm_mul_pd(velec,rinvsq21);
1318
1319             fscal            = felec;
1320
1321             /* Update vectorial force */
1322             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1323             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1324             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1325
1326             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1327             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1328             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1329
1330             /**************************
1331              * CALCULATE INTERACTIONS *
1332              **************************/
1333
1334             /* COULOMB ELECTROSTATICS */
1335             velec            = _mm_mul_pd(qq22,rinv22);
1336             felec            = _mm_mul_pd(velec,rinvsq22);
1337
1338             fscal            = felec;
1339
1340             /* Update vectorial force */
1341             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1342             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1343             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1344
1345             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1346             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1347             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1348
1349             /**************************
1350              * CALCULATE INTERACTIONS *
1351              **************************/
1352
1353             /* COULOMB ELECTROSTATICS */
1354             velec            = _mm_mul_pd(qq23,rinv23);
1355             felec            = _mm_mul_pd(velec,rinvsq23);
1356
1357             fscal            = felec;
1358
1359             /* Update vectorial force */
1360             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1361             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1362             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1363
1364             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1365             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1366             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1367
1368             /**************************
1369              * CALCULATE INTERACTIONS *
1370              **************************/
1371
1372             /* COULOMB ELECTROSTATICS */
1373             velec            = _mm_mul_pd(qq31,rinv31);
1374             felec            = _mm_mul_pd(velec,rinvsq31);
1375
1376             fscal            = felec;
1377
1378             /* Update vectorial force */
1379             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1380             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1381             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1382
1383             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1384             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1385             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1386
1387             /**************************
1388              * CALCULATE INTERACTIONS *
1389              **************************/
1390
1391             /* COULOMB ELECTROSTATICS */
1392             velec            = _mm_mul_pd(qq32,rinv32);
1393             felec            = _mm_mul_pd(velec,rinvsq32);
1394
1395             fscal            = felec;
1396
1397             /* Update vectorial force */
1398             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1399             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1400             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1401
1402             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1403             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1404             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1405
1406             /**************************
1407              * CALCULATE INTERACTIONS *
1408              **************************/
1409
1410             /* COULOMB ELECTROSTATICS */
1411             velec            = _mm_mul_pd(qq33,rinv33);
1412             felec            = _mm_mul_pd(velec,rinvsq33);
1413
1414             fscal            = felec;
1415
1416             /* Update vectorial force */
1417             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1418             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1419             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1420
1421             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1422             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1423             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1424
1425             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);
1426
1427             /* Inner loop uses 324 flops */
1428         }
1429
1430         if(jidx<j_index_end)
1431         {
1432
1433             jnrA             = jjnr[jidx];
1434             j_coord_offsetA  = DIM*jnrA;
1435
1436             /* load j atom coordinates */
1437             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1438                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1439                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1440
1441             /* Calculate displacement vector */
1442             dx00             = _mm_sub_pd(ix0,jx0);
1443             dy00             = _mm_sub_pd(iy0,jy0);
1444             dz00             = _mm_sub_pd(iz0,jz0);
1445             dx11             = _mm_sub_pd(ix1,jx1);
1446             dy11             = _mm_sub_pd(iy1,jy1);
1447             dz11             = _mm_sub_pd(iz1,jz1);
1448             dx12             = _mm_sub_pd(ix1,jx2);
1449             dy12             = _mm_sub_pd(iy1,jy2);
1450             dz12             = _mm_sub_pd(iz1,jz2);
1451             dx13             = _mm_sub_pd(ix1,jx3);
1452             dy13             = _mm_sub_pd(iy1,jy3);
1453             dz13             = _mm_sub_pd(iz1,jz3);
1454             dx21             = _mm_sub_pd(ix2,jx1);
1455             dy21             = _mm_sub_pd(iy2,jy1);
1456             dz21             = _mm_sub_pd(iz2,jz1);
1457             dx22             = _mm_sub_pd(ix2,jx2);
1458             dy22             = _mm_sub_pd(iy2,jy2);
1459             dz22             = _mm_sub_pd(iz2,jz2);
1460             dx23             = _mm_sub_pd(ix2,jx3);
1461             dy23             = _mm_sub_pd(iy2,jy3);
1462             dz23             = _mm_sub_pd(iz2,jz3);
1463             dx31             = _mm_sub_pd(ix3,jx1);
1464             dy31             = _mm_sub_pd(iy3,jy1);
1465             dz31             = _mm_sub_pd(iz3,jz1);
1466             dx32             = _mm_sub_pd(ix3,jx2);
1467             dy32             = _mm_sub_pd(iy3,jy2);
1468             dz32             = _mm_sub_pd(iz3,jz2);
1469             dx33             = _mm_sub_pd(ix3,jx3);
1470             dy33             = _mm_sub_pd(iy3,jy3);
1471             dz33             = _mm_sub_pd(iz3,jz3);
1472
1473             /* Calculate squared distance and things based on it */
1474             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1475             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1476             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1477             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1478             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1479             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1480             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1481             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1482             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1483             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1484
1485             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1486             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1487             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1488             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1489             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1490             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1491             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1492             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1493             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1494             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1495
1496             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1497             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1498             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1499             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1500             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1501             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1502             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1503             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1504             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1505
1506             fjx0             = _mm_setzero_pd();
1507             fjy0             = _mm_setzero_pd();
1508             fjz0             = _mm_setzero_pd();
1509             fjx1             = _mm_setzero_pd();
1510             fjy1             = _mm_setzero_pd();
1511             fjz1             = _mm_setzero_pd();
1512             fjx2             = _mm_setzero_pd();
1513             fjy2             = _mm_setzero_pd();
1514             fjz2             = _mm_setzero_pd();
1515             fjx3             = _mm_setzero_pd();
1516             fjy3             = _mm_setzero_pd();
1517             fjz3             = _mm_setzero_pd();
1518
1519             /**************************
1520              * CALCULATE INTERACTIONS *
1521              **************************/
1522
1523             r00              = _mm_mul_pd(rsq00,rinv00);
1524
1525             /* Calculate table index by multiplying r with table scale and truncate to integer */
1526             rt               = _mm_mul_pd(r00,vftabscale);
1527             vfitab           = _mm_cvttpd_epi32(rt);
1528 #ifdef __XOP__
1529             vfeps            = _mm_frcz_pd(rt);
1530 #else
1531             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
1532 #endif
1533             twovfeps         = _mm_add_pd(vfeps,vfeps);
1534             vfitab           = _mm_slli_epi32(vfitab,3);
1535
1536             /* CUBIC SPLINE TABLE DISPERSION */
1537             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1538             F                = _mm_setzero_pd();
1539             GMX_MM_TRANSPOSE2_PD(Y,F);
1540             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1541             H                = _mm_setzero_pd();
1542             GMX_MM_TRANSPOSE2_PD(G,H);
1543             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1544             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1545             fvdw6            = _mm_mul_pd(c6_00,FF);
1546
1547             /* CUBIC SPLINE TABLE REPULSION */
1548             vfitab           = _mm_add_epi32(vfitab,ifour);
1549             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1550             F                = _mm_setzero_pd();
1551             GMX_MM_TRANSPOSE2_PD(Y,F);
1552             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1553             H                = _mm_setzero_pd();
1554             GMX_MM_TRANSPOSE2_PD(G,H);
1555             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1556             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1557             fvdw12           = _mm_mul_pd(c12_00,FF);
1558             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1559
1560             fscal            = fvdw;
1561
1562             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1563
1564             /* Update vectorial force */
1565             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1566             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1567             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1568
1569             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1570             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1571             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1572
1573             /**************************
1574              * CALCULATE INTERACTIONS *
1575              **************************/
1576
1577             /* COULOMB ELECTROSTATICS */
1578             velec            = _mm_mul_pd(qq11,rinv11);
1579             felec            = _mm_mul_pd(velec,rinvsq11);
1580
1581             fscal            = felec;
1582
1583             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1584
1585             /* Update vectorial force */
1586             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1587             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1588             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1589
1590             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1591             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1592             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1593
1594             /**************************
1595              * CALCULATE INTERACTIONS *
1596              **************************/
1597
1598             /* COULOMB ELECTROSTATICS */
1599             velec            = _mm_mul_pd(qq12,rinv12);
1600             felec            = _mm_mul_pd(velec,rinvsq12);
1601
1602             fscal            = felec;
1603
1604             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1605
1606             /* Update vectorial force */
1607             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1608             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1609             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1610
1611             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1612             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1613             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1614
1615             /**************************
1616              * CALCULATE INTERACTIONS *
1617              **************************/
1618
1619             /* COULOMB ELECTROSTATICS */
1620             velec            = _mm_mul_pd(qq13,rinv13);
1621             felec            = _mm_mul_pd(velec,rinvsq13);
1622
1623             fscal            = felec;
1624
1625             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1626
1627             /* Update vectorial force */
1628             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1629             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1630             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1631
1632             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1633             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1634             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1635
1636             /**************************
1637              * CALCULATE INTERACTIONS *
1638              **************************/
1639
1640             /* COULOMB ELECTROSTATICS */
1641             velec            = _mm_mul_pd(qq21,rinv21);
1642             felec            = _mm_mul_pd(velec,rinvsq21);
1643
1644             fscal            = felec;
1645
1646             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1647
1648             /* Update vectorial force */
1649             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1650             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1651             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1652
1653             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1654             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1655             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1656
1657             /**************************
1658              * CALCULATE INTERACTIONS *
1659              **************************/
1660
1661             /* COULOMB ELECTROSTATICS */
1662             velec            = _mm_mul_pd(qq22,rinv22);
1663             felec            = _mm_mul_pd(velec,rinvsq22);
1664
1665             fscal            = felec;
1666
1667             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1668
1669             /* Update vectorial force */
1670             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1671             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1672             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1673
1674             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1675             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1676             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1677
1678             /**************************
1679              * CALCULATE INTERACTIONS *
1680              **************************/
1681
1682             /* COULOMB ELECTROSTATICS */
1683             velec            = _mm_mul_pd(qq23,rinv23);
1684             felec            = _mm_mul_pd(velec,rinvsq23);
1685
1686             fscal            = felec;
1687
1688             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1689
1690             /* Update vectorial force */
1691             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1692             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1693             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1694
1695             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1696             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1697             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1698
1699             /**************************
1700              * CALCULATE INTERACTIONS *
1701              **************************/
1702
1703             /* COULOMB ELECTROSTATICS */
1704             velec            = _mm_mul_pd(qq31,rinv31);
1705             felec            = _mm_mul_pd(velec,rinvsq31);
1706
1707             fscal            = felec;
1708
1709             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1710
1711             /* Update vectorial force */
1712             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1713             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1714             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1715
1716             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1717             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1718             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1719
1720             /**************************
1721              * CALCULATE INTERACTIONS *
1722              **************************/
1723
1724             /* COULOMB ELECTROSTATICS */
1725             velec            = _mm_mul_pd(qq32,rinv32);
1726             felec            = _mm_mul_pd(velec,rinvsq32);
1727
1728             fscal            = felec;
1729
1730             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1731
1732             /* Update vectorial force */
1733             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1734             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1735             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1736
1737             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1738             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1739             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1740
1741             /**************************
1742              * CALCULATE INTERACTIONS *
1743              **************************/
1744
1745             /* COULOMB ELECTROSTATICS */
1746             velec            = _mm_mul_pd(qq33,rinv33);
1747             felec            = _mm_mul_pd(velec,rinvsq33);
1748
1749             fscal            = felec;
1750
1751             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1752
1753             /* Update vectorial force */
1754             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1755             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1756             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1757
1758             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1759             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1760             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1761
1762             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1763
1764             /* Inner loop uses 324 flops */
1765         }
1766
1767         /* End of innermost loop */
1768
1769         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1770                                               f+i_coord_offset,fshift+i_shift_offset);
1771
1772         /* Increment number of inner iterations */
1773         inneriter                  += j_index_end - j_index_start;
1774
1775         /* Outer loop uses 24 flops */
1776     }
1777
1778     /* Increment number of outer iterations */
1779     outeriter        += nri;
1780
1781     /* Update outer/inner flops */
1782
1783     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*324);
1784 }