src/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecRFCut_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_ElecRFCut_VdwCSTab_GeomW4W4_VF_avx_128_fma_double
  38  * Electrostatics interaction: ReactionField
  39  * VdW interaction:            CubicSplineTable
  40  * Geometry:                   Water4-Water4
  41  * Calculate force/pot:        PotentialAndForce
  42  */
  43 void
  44 nb_kernel_ElecRFCut_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     krf              = _mm_set1_pd(fr->ic->k_rf);
 122     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 123     crf              = _mm_set1_pd(fr->ic->c_rf);
 124     nvdwtype         = fr->ntype;
 125     vdwparam         = fr->nbfp;
 126     vdwtype          = mdatoms->typeA;
 127
 128     vftab            = kernel_data->table_vdw->data;
 129     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
 130
 131     /* Setup water-specific parameters */
 132     inr              = nlist->iinr[0];
 133     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 134     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 135     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 136     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 137
 138     jq1              = _mm_set1_pd(charge[inr+1]);
 139     jq2              = _mm_set1_pd(charge[inr+2]);
 140     jq3              = _mm_set1_pd(charge[inr+3]);
 141     vdwjidx0A        = 2*vdwtype[inr+0];
 142     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 143     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 144     qq11             = _mm_mul_pd(iq1,jq1);
 145     qq12             = _mm_mul_pd(iq1,jq2);
 146     qq13             = _mm_mul_pd(iq1,jq3);
 147     qq21             = _mm_mul_pd(iq2,jq1);
 148     qq22             = _mm_mul_pd(iq2,jq2);
 149     qq23             = _mm_mul_pd(iq2,jq3);
 150     qq31             = _mm_mul_pd(iq3,jq1);
 151     qq32             = _mm_mul_pd(iq3,jq2);
 152     qq33             = _mm_mul_pd(iq3,jq3);
 153
 154     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 155     rcutoff_scalar   = fr->rcoulomb;
 156     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 157     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 158
 159     /* Avoid stupid compiler warnings */
 160     jnrA = jnrB = 0;
 161     j_coord_offsetA = 0;
 162     j_coord_offsetB = 0;
 163
 164     outeriter        = 0;
 165     inneriter        = 0;
 166
 167     /* Start outer loop over neighborlists */
 168     for(iidx=0; iidx<nri; iidx++)
 169     {
 170         /* Load shift vector for this list */
 171         i_shift_offset   = DIM*shiftidx[iidx];
 172
 173         /* Load limits for loop over neighbors */
 174         j_index_start    = jindex[iidx];
 175         j_index_end      = jindex[iidx+1];
 176
 177         /* Get outer coordinate index */
 178         inr              = iinr[iidx];
 179         i_coord_offset   = DIM*inr;
 180
 181         /* Load i particle coords and add shift vector */
 182         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 183                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 184
 185         fix0             = _mm_setzero_pd();
 186         fiy0             = _mm_setzero_pd();
 187         fiz0             = _mm_setzero_pd();
 188         fix1             = _mm_setzero_pd();
 189         fiy1             = _mm_setzero_pd();
 190         fiz1             = _mm_setzero_pd();
 191         fix2             = _mm_setzero_pd();
 192         fiy2             = _mm_setzero_pd();
 193         fiz2             = _mm_setzero_pd();
 194         fix3             = _mm_setzero_pd();
 195         fiy3             = _mm_setzero_pd();
 196         fiz3             = _mm_setzero_pd();
 197
 198         /* Reset potential sums */
 199         velecsum         = _mm_setzero_pd();
 200         vvdwsum          = _mm_setzero_pd();
 201
 202         /* Start inner kernel loop */
 203         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 204         {
 205
 206             /* Get j neighbor index, and coordinate index */
 207             jnrA             = jjnr[jidx];
 208             jnrB             = jjnr[jidx+1];
 209             j_coord_offsetA  = DIM*jnrA;
 210             j_coord_offsetB  = DIM*jnrB;
 211
 212             /* load j atom coordinates */
 213             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 214                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 215                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 216
 217             /* Calculate displacement vector */
 218             dx00             = _mm_sub_pd(ix0,jx0);
 219             dy00             = _mm_sub_pd(iy0,jy0);
 220             dz00             = _mm_sub_pd(iz0,jz0);
 221             dx11             = _mm_sub_pd(ix1,jx1);
 222             dy11             = _mm_sub_pd(iy1,jy1);
 223             dz11             = _mm_sub_pd(iz1,jz1);
 224             dx12             = _mm_sub_pd(ix1,jx2);
 225             dy12             = _mm_sub_pd(iy1,jy2);
 226             dz12             = _mm_sub_pd(iz1,jz2);
 227             dx13             = _mm_sub_pd(ix1,jx3);
 228             dy13             = _mm_sub_pd(iy1,jy3);
 229             dz13             = _mm_sub_pd(iz1,jz3);
 230             dx21             = _mm_sub_pd(ix2,jx1);
 231             dy21             = _mm_sub_pd(iy2,jy1);
 232             dz21             = _mm_sub_pd(iz2,jz1);
 233             dx22             = _mm_sub_pd(ix2,jx2);
 234             dy22             = _mm_sub_pd(iy2,jy2);
 235             dz22             = _mm_sub_pd(iz2,jz2);
 236             dx23             = _mm_sub_pd(ix2,jx3);
 237             dy23             = _mm_sub_pd(iy2,jy3);
 238             dz23             = _mm_sub_pd(iz2,jz3);
 239             dx31             = _mm_sub_pd(ix3,jx1);
 240             dy31             = _mm_sub_pd(iy3,jy1);
 241             dz31             = _mm_sub_pd(iz3,jz1);
 242             dx32             = _mm_sub_pd(ix3,jx2);
 243             dy32             = _mm_sub_pd(iy3,jy2);
 244             dz32             = _mm_sub_pd(iz3,jz2);
 245             dx33             = _mm_sub_pd(ix3,jx3);
 246             dy33             = _mm_sub_pd(iy3,jy3);
 247             dz33             = _mm_sub_pd(iz3,jz3);
 248
 249             /* Calculate squared distance and things based on it */
 250             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 251             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 252             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 253             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 254             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 255             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 256             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 257             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 258             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 259             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 260
 261             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 262             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 263             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 264             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 265             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 266             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 267             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 268             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 269             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 270             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 271
 272             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 273             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 274             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 275             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 276             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 277             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 278             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 279             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 280             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 281
 282             fjx0             = _mm_setzero_pd();
 283             fjy0             = _mm_setzero_pd();
 284             fjz0             = _mm_setzero_pd();
 285             fjx1             = _mm_setzero_pd();
 286             fjy1             = _mm_setzero_pd();
 287             fjz1             = _mm_setzero_pd();
 288             fjx2             = _mm_setzero_pd();
 289             fjy2             = _mm_setzero_pd();
 290             fjz2             = _mm_setzero_pd();
 291             fjx3             = _mm_setzero_pd();
 292             fjy3             = _mm_setzero_pd();
 293             fjz3             = _mm_setzero_pd();
 294
 295             /**************************
 296              * CALCULATE INTERACTIONS *
 297              **************************/
 298
 299             r00              = _mm_mul_pd(rsq00,rinv00);
 300
 301             /* Calculate table index by multiplying r with table scale and truncate to integer */
 302             rt               = _mm_mul_pd(r00,vftabscale);
 303             vfitab           = _mm_cvttpd_epi32(rt);
 304 #ifdef __XOP__
 305             vfeps            = _mm_frcz_pd(rt);
 306 #else
 307             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 308 #endif
 309             twovfeps         = _mm_add_pd(vfeps,vfeps);
 310             vfitab           = _mm_slli_epi32(vfitab,3);
 311
 312             /* CUBIC SPLINE TABLE DISPERSION */
 313             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 314             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 315             GMX_MM_TRANSPOSE2_PD(Y,F);
 316             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 317             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
 318             GMX_MM_TRANSPOSE2_PD(G,H);
 319             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 320             VV               = _mm_macc_pd(vfeps,Fp,Y);
 321             vvdw6            = _mm_mul_pd(c6_00,VV);
 322             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 323             fvdw6            = _mm_mul_pd(c6_00,FF);
 324
 325             /* CUBIC SPLINE TABLE REPULSION */
 326             vfitab           = _mm_add_epi32(vfitab,ifour);
 327             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 328             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 329             GMX_MM_TRANSPOSE2_PD(Y,F);
 330             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 331             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
 332             GMX_MM_TRANSPOSE2_PD(G,H);
 333             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 334             VV               = _mm_macc_pd(vfeps,Fp,Y);
 335             vvdw12           = _mm_mul_pd(c12_00,VV);
 336             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 337             fvdw12           = _mm_mul_pd(c12_00,FF);
 338             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 339             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 340
 341             /* Update potential sum for this i atom from the interaction with this j atom. */
 342             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 343
 344             fscal            = fvdw;
 345
 346             /* Update vectorial force */
 347             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 348             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 349             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 350
 351             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 352             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 353             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 354
 355             /**************************
 356              * CALCULATE INTERACTIONS *
 357              **************************/
 358
 359             if (gmx_mm_any_lt(rsq11,rcutoff2))
 360             {
 361
 362             /* REACTION-FIELD ELECTROSTATICS */
 363             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 364             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 365
 366             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 367
 368             /* Update potential sum for this i atom from the interaction with this j atom. */
 369             velec            = _mm_and_pd(velec,cutoff_mask);
 370             velecsum         = _mm_add_pd(velecsum,velec);
 371
 372             fscal            = felec;
 373
 374             fscal            = _mm_and_pd(fscal,cutoff_mask);
 375
 376             /* Update vectorial force */
 377             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 378             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 379             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 380
 381             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 382             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 383             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 384
 385             }
 386
 387             /**************************
 388              * CALCULATE INTERACTIONS *
 389              **************************/
 390
 391             if (gmx_mm_any_lt(rsq12,rcutoff2))
 392             {
 393
 394             /* REACTION-FIELD ELECTROSTATICS */
 395             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 396             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 397
 398             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 399
 400             /* Update potential sum for this i atom from the interaction with this j atom. */
 401             velec            = _mm_and_pd(velec,cutoff_mask);
 402             velecsum         = _mm_add_pd(velecsum,velec);
 403
 404             fscal            = felec;
 405
 406             fscal            = _mm_and_pd(fscal,cutoff_mask);
 407
 408             /* Update vectorial force */
 409             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 410             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 411             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 412
 413             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 414             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 415             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 416
 417             }
 418
 419             /**************************
 420              * CALCULATE INTERACTIONS *
 421              **************************/
 422
 423             if (gmx_mm_any_lt(rsq13,rcutoff2))
 424             {
 425
 426             /* REACTION-FIELD ELECTROSTATICS */
 427             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 428             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 429
 430             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 431
 432             /* Update potential sum for this i atom from the interaction with this j atom. */
 433             velec            = _mm_and_pd(velec,cutoff_mask);
 434             velecsum         = _mm_add_pd(velecsum,velec);
 435
 436             fscal            = felec;
 437
 438             fscal            = _mm_and_pd(fscal,cutoff_mask);
 439
 440             /* Update vectorial force */
 441             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 442             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 443             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 444
 445             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 446             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 447             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 448
 449             }
 450
 451             /**************************
 452              * CALCULATE INTERACTIONS *
 453              **************************/
 454
 455             if (gmx_mm_any_lt(rsq21,rcutoff2))
 456             {
 457
 458             /* REACTION-FIELD ELECTROSTATICS */
 459             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 460             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 461
 462             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 463
 464             /* Update potential sum for this i atom from the interaction with this j atom. */
 465             velec            = _mm_and_pd(velec,cutoff_mask);
 466             velecsum         = _mm_add_pd(velecsum,velec);
 467
 468             fscal            = felec;
 469
 470             fscal            = _mm_and_pd(fscal,cutoff_mask);
 471
 472             /* Update vectorial force */
 473             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 474             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 475             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 476
 477             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 478             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 479             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 480
 481             }
 482
 483             /**************************
 484              * CALCULATE INTERACTIONS *
 485              **************************/
 486
 487             if (gmx_mm_any_lt(rsq22,rcutoff2))
 488             {
 489
 490             /* REACTION-FIELD ELECTROSTATICS */
 491             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 492             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 493
 494             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 495
 496             /* Update potential sum for this i atom from the interaction with this j atom. */
 497             velec            = _mm_and_pd(velec,cutoff_mask);
 498             velecsum         = _mm_add_pd(velecsum,velec);
 499
 500             fscal            = felec;
 501
 502             fscal            = _mm_and_pd(fscal,cutoff_mask);
 503
 504             /* Update vectorial force */
 505             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 506             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 507             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 508
 509             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 510             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 511             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 512
 513             }
 514
 515             /**************************
 516              * CALCULATE INTERACTIONS *
 517              **************************/
 518
 519             if (gmx_mm_any_lt(rsq23,rcutoff2))
 520             {
 521
 522             /* REACTION-FIELD ELECTROSTATICS */
 523             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 524             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 525
 526             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 527
 528             /* Update potential sum for this i atom from the interaction with this j atom. */
 529             velec            = _mm_and_pd(velec,cutoff_mask);
 530             velecsum         = _mm_add_pd(velecsum,velec);
 531
 532             fscal            = felec;
 533
 534             fscal            = _mm_and_pd(fscal,cutoff_mask);
 535
 536             /* Update vectorial force */
 537             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 538             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 539             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 540
 541             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 542             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 543             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 544
 545             }
 546
 547             /**************************
 548              * CALCULATE INTERACTIONS *
 549              **************************/
 550
 551             if (gmx_mm_any_lt(rsq31,rcutoff2))
 552             {
 553
 554             /* REACTION-FIELD ELECTROSTATICS */
 555             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
 556             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
 557
 558             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
 559
 560             /* Update potential sum for this i atom from the interaction with this j atom. */
 561             velec            = _mm_and_pd(velec,cutoff_mask);
 562             velecsum         = _mm_add_pd(velecsum,velec);
 563
 564             fscal            = felec;
 565
 566             fscal            = _mm_and_pd(fscal,cutoff_mask);
 567
 568             /* Update vectorial force */
 569             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 570             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 571             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 572
 573             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 574             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 575             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 576
 577             }
 578
 579             /**************************
 580              * CALCULATE INTERACTIONS *
 581              **************************/
 582
 583             if (gmx_mm_any_lt(rsq32,rcutoff2))
 584             {
 585
 586             /* REACTION-FIELD ELECTROSTATICS */
 587             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
 588             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
 589
 590             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
 591
 592             /* Update potential sum for this i atom from the interaction with this j atom. */
 593             velec            = _mm_and_pd(velec,cutoff_mask);
 594             velecsum         = _mm_add_pd(velecsum,velec);
 595
 596             fscal            = felec;
 597
 598             fscal            = _mm_and_pd(fscal,cutoff_mask);
 599
 600             /* Update vectorial force */
 601             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 602             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 603             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 604
 605             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 606             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 607             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 608
 609             }
 610
 611             /**************************
 612              * CALCULATE INTERACTIONS *
 613              **************************/
 614
 615             if (gmx_mm_any_lt(rsq33,rcutoff2))
 616             {
 617
 618             /* REACTION-FIELD ELECTROSTATICS */
 619             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
 620             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
 621
 622             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
 623
 624             /* Update potential sum for this i atom from the interaction with this j atom. */
 625             velec            = _mm_and_pd(velec,cutoff_mask);
 626             velecsum         = _mm_add_pd(velecsum,velec);
 627
 628             fscal            = felec;
 629
 630             fscal            = _mm_and_pd(fscal,cutoff_mask);
 631
 632             /* Update vectorial force */
 633             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 634             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 635             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 636
 637             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 638             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 639             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 640
 641             }
 642
 643             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);
 644
 645             /* Inner loop uses 413 flops */
 646         }
 647
 648         if(jidx<j_index_end)
 649         {
 650
 651             jnrA             = jjnr[jidx];
 652             j_coord_offsetA  = DIM*jnrA;
 653
 654             /* load j atom coordinates */
 655             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 656                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 657                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 658
 659             /* Calculate displacement vector */
 660             dx00             = _mm_sub_pd(ix0,jx0);
 661             dy00             = _mm_sub_pd(iy0,jy0);
 662             dz00             = _mm_sub_pd(iz0,jz0);
 663             dx11             = _mm_sub_pd(ix1,jx1);
 664             dy11             = _mm_sub_pd(iy1,jy1);
 665             dz11             = _mm_sub_pd(iz1,jz1);
 666             dx12             = _mm_sub_pd(ix1,jx2);
 667             dy12             = _mm_sub_pd(iy1,jy2);
 668             dz12             = _mm_sub_pd(iz1,jz2);
 669             dx13             = _mm_sub_pd(ix1,jx3);
 670             dy13             = _mm_sub_pd(iy1,jy3);
 671             dz13             = _mm_sub_pd(iz1,jz3);
 672             dx21             = _mm_sub_pd(ix2,jx1);
 673             dy21             = _mm_sub_pd(iy2,jy1);
 674             dz21             = _mm_sub_pd(iz2,jz1);
 675             dx22             = _mm_sub_pd(ix2,jx2);
 676             dy22             = _mm_sub_pd(iy2,jy2);
 677             dz22             = _mm_sub_pd(iz2,jz2);
 678             dx23             = _mm_sub_pd(ix2,jx3);
 679             dy23             = _mm_sub_pd(iy2,jy3);
 680             dz23             = _mm_sub_pd(iz2,jz3);
 681             dx31             = _mm_sub_pd(ix3,jx1);
 682             dy31             = _mm_sub_pd(iy3,jy1);
 683             dz31             = _mm_sub_pd(iz3,jz1);
 684             dx32             = _mm_sub_pd(ix3,jx2);
 685             dy32             = _mm_sub_pd(iy3,jy2);
 686             dz32             = _mm_sub_pd(iz3,jz2);
 687             dx33             = _mm_sub_pd(ix3,jx3);
 688             dy33             = _mm_sub_pd(iy3,jy3);
 689             dz33             = _mm_sub_pd(iz3,jz3);
 690
 691             /* Calculate squared distance and things based on it */
 692             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 693             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 694             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 695             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 696             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 697             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 698             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 699             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 700             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 701             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 702
 703             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 704             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 705             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 706             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 707             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 708             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 709             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 710             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 711             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 712             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 713
 714             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 715             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 716             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 717             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 718             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 719             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 720             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 721             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 722             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 723
 724             fjx0             = _mm_setzero_pd();
 725             fjy0             = _mm_setzero_pd();
 726             fjz0             = _mm_setzero_pd();
 727             fjx1             = _mm_setzero_pd();
 728             fjy1             = _mm_setzero_pd();
 729             fjz1             = _mm_setzero_pd();
 730             fjx2             = _mm_setzero_pd();
 731             fjy2             = _mm_setzero_pd();
 732             fjz2             = _mm_setzero_pd();
 733             fjx3             = _mm_setzero_pd();
 734             fjy3             = _mm_setzero_pd();
 735             fjz3             = _mm_setzero_pd();
 736
 737             /**************************
 738              * CALCULATE INTERACTIONS *
 739              **************************/
 740
 741             r00              = _mm_mul_pd(rsq00,rinv00);
 742
 743             /* Calculate table index by multiplying r with table scale and truncate to integer */
 744             rt               = _mm_mul_pd(r00,vftabscale);
 745             vfitab           = _mm_cvttpd_epi32(rt);
 746 #ifdef __XOP__
 747             vfeps            = _mm_frcz_pd(rt);
 748 #else
 749             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 750 #endif
 751             twovfeps         = _mm_add_pd(vfeps,vfeps);
 752             vfitab           = _mm_slli_epi32(vfitab,3);
 753
 754             /* CUBIC SPLINE TABLE DISPERSION */
 755             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 756             F                = _mm_setzero_pd();
 757             GMX_MM_TRANSPOSE2_PD(Y,F);
 758             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 759             H                = _mm_setzero_pd();
 760             GMX_MM_TRANSPOSE2_PD(G,H);
 761             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 762             VV               = _mm_macc_pd(vfeps,Fp,Y);
 763             vvdw6            = _mm_mul_pd(c6_00,VV);
 764             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 765             fvdw6            = _mm_mul_pd(c6_00,FF);
 766
 767             /* CUBIC SPLINE TABLE REPULSION */
 768             vfitab           = _mm_add_epi32(vfitab,ifour);
 769             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 770             F                = _mm_setzero_pd();
 771             GMX_MM_TRANSPOSE2_PD(Y,F);
 772             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 773             H                = _mm_setzero_pd();
 774             GMX_MM_TRANSPOSE2_PD(G,H);
 775             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 776             VV               = _mm_macc_pd(vfeps,Fp,Y);
 777             vvdw12           = _mm_mul_pd(c12_00,VV);
 778             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 779             fvdw12           = _mm_mul_pd(c12_00,FF);
 780             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 781             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 782
 783             /* Update potential sum for this i atom from the interaction with this j atom. */
 784             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 785             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 786
 787             fscal            = fvdw;
 788
 789             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 790
 791             /* Update vectorial force */
 792             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 793             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 794             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 795
 796             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 797             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 798             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 799
 800             /**************************
 801              * CALCULATE INTERACTIONS *
 802              **************************/
 803
 804             if (gmx_mm_any_lt(rsq11,rcutoff2))
 805             {
 806
 807             /* REACTION-FIELD ELECTROSTATICS */
 808             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 809             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 810
 811             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 812
 813             /* Update potential sum for this i atom from the interaction with this j atom. */
 814             velec            = _mm_and_pd(velec,cutoff_mask);
 815             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 816             velecsum         = _mm_add_pd(velecsum,velec);
 817
 818             fscal            = felec;
 819
 820             fscal            = _mm_and_pd(fscal,cutoff_mask);
 821
 822             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 823
 824             /* Update vectorial force */
 825             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 826             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 827             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 828
 829             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 830             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 831             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 832
 833             }
 834
 835             /**************************
 836              * CALCULATE INTERACTIONS *
 837              **************************/
 838
 839             if (gmx_mm_any_lt(rsq12,rcutoff2))
 840             {
 841
 842             /* REACTION-FIELD ELECTROSTATICS */
 843             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 844             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 845
 846             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 847
 848             /* Update potential sum for this i atom from the interaction with this j atom. */
 849             velec            = _mm_and_pd(velec,cutoff_mask);
 850             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 851             velecsum         = _mm_add_pd(velecsum,velec);
 852
 853             fscal            = felec;
 854
 855             fscal            = _mm_and_pd(fscal,cutoff_mask);
 856
 857             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 858
 859             /* Update vectorial force */
 860             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 861             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 862             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 863
 864             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 865             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 866             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 867
 868             }
 869
 870             /**************************
 871              * CALCULATE INTERACTIONS *
 872              **************************/
 873
 874             if (gmx_mm_any_lt(rsq13,rcutoff2))
 875             {
 876
 877             /* REACTION-FIELD ELECTROSTATICS */
 878             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 879             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 880
 881             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
 882
 883             /* Update potential sum for this i atom from the interaction with this j atom. */
 884             velec            = _mm_and_pd(velec,cutoff_mask);
 885             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 886             velecsum         = _mm_add_pd(velecsum,velec);
 887
 888             fscal            = felec;
 889
 890             fscal            = _mm_and_pd(fscal,cutoff_mask);
 891
 892             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 893
 894             /* Update vectorial force */
 895             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 896             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 897             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 898
 899             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 900             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 901             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 902
 903             }
 904
 905             /**************************
 906              * CALCULATE INTERACTIONS *
 907              **************************/
 908
 909             if (gmx_mm_any_lt(rsq21,rcutoff2))
 910             {
 911
 912             /* REACTION-FIELD ELECTROSTATICS */
 913             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 914             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 915
 916             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 917
 918             /* Update potential sum for this i atom from the interaction with this j atom. */
 919             velec            = _mm_and_pd(velec,cutoff_mask);
 920             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 921             velecsum         = _mm_add_pd(velecsum,velec);
 922
 923             fscal            = felec;
 924
 925             fscal            = _mm_and_pd(fscal,cutoff_mask);
 926
 927             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 928
 929             /* Update vectorial force */
 930             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 931             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 932             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 933
 934             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 935             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 936             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 937
 938             }
 939
 940             /**************************
 941              * CALCULATE INTERACTIONS *
 942              **************************/
 943
 944             if (gmx_mm_any_lt(rsq22,rcutoff2))
 945             {
 946
 947             /* REACTION-FIELD ELECTROSTATICS */
 948             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 949             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 950
 951             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 952
 953             /* Update potential sum for this i atom from the interaction with this j atom. */
 954             velec            = _mm_and_pd(velec,cutoff_mask);
 955             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 956             velecsum         = _mm_add_pd(velecsum,velec);
 957
 958             fscal            = felec;
 959
 960             fscal            = _mm_and_pd(fscal,cutoff_mask);
 961
 962             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 963
 964             /* Update vectorial force */
 965             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 966             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 967             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 968
 969             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 970             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 971             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 972
 973             }
 974
 975             /**************************
 976              * CALCULATE INTERACTIONS *
 977              **************************/
 978
 979             if (gmx_mm_any_lt(rsq23,rcutoff2))
 980             {
 981
 982             /* REACTION-FIELD ELECTROSTATICS */
 983             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 984             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 985
 986             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
 987
 988             /* Update potential sum for this i atom from the interaction with this j atom. */
 989             velec            = _mm_and_pd(velec,cutoff_mask);
 990             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 991             velecsum         = _mm_add_pd(velecsum,velec);
 992
 993             fscal            = felec;
 994
 995             fscal            = _mm_and_pd(fscal,cutoff_mask);
 996
 997             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 998
 999             /* Update vectorial force */
1000             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1001             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1002             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1003
1004             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1005             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1006             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1007
1008             }
1009
1010             /**************************
1011              * CALCULATE INTERACTIONS *
1012              **************************/
1013
1014             if (gmx_mm_any_lt(rsq31,rcutoff2))
1015             {
1016
1017             /* REACTION-FIELD ELECTROSTATICS */
1018             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
1019             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1020
1021             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1022
1023             /* Update potential sum for this i atom from the interaction with this j atom. */
1024             velec            = _mm_and_pd(velec,cutoff_mask);
1025             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1026             velecsum         = _mm_add_pd(velecsum,velec);
1027
1028             fscal            = felec;
1029
1030             fscal            = _mm_and_pd(fscal,cutoff_mask);
1031
1032             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1033
1034             /* Update vectorial force */
1035             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1036             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1037             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1038
1039             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1040             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1041             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1042
1043             }
1044
1045             /**************************
1046              * CALCULATE INTERACTIONS *
1047              **************************/
1048
1049             if (gmx_mm_any_lt(rsq32,rcutoff2))
1050             {
1051
1052             /* REACTION-FIELD ELECTROSTATICS */
1053             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
1054             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1055
1056             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1057
1058             /* Update potential sum for this i atom from the interaction with this j atom. */
1059             velec            = _mm_and_pd(velec,cutoff_mask);
1060             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1061             velecsum         = _mm_add_pd(velecsum,velec);
1062
1063             fscal            = felec;
1064
1065             fscal            = _mm_and_pd(fscal,cutoff_mask);
1066
1067             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1068
1069             /* Update vectorial force */
1070             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1071             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1072             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1073
1074             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1075             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1076             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1077
1078             }
1079
1080             /**************************
1081              * CALCULATE INTERACTIONS *
1082              **************************/
1083
1084             if (gmx_mm_any_lt(rsq33,rcutoff2))
1085             {
1086
1087             /* REACTION-FIELD ELECTROSTATICS */
1088             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
1089             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1090
1091             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1092
1093             /* Update potential sum for this i atom from the interaction with this j atom. */
1094             velec            = _mm_and_pd(velec,cutoff_mask);
1095             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1096             velecsum         = _mm_add_pd(velecsum,velec);
1097
1098             fscal            = felec;
1099
1100             fscal            = _mm_and_pd(fscal,cutoff_mask);
1101
1102             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1103
1104             /* Update vectorial force */
1105             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1106             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1107             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1108
1109             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1110             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1111             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1112
1113             }
1114
1115             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1116
1117             /* Inner loop uses 413 flops */
1118         }
1119
1120         /* End of innermost loop */
1121
1122         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1123                                               f+i_coord_offset,fshift+i_shift_offset);
1124
1125         ggid                        = gid[iidx];
1126         /* Update potential energies */
1127         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1128         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
1129
1130         /* Increment number of inner iterations */
1131         inneriter                  += j_index_end - j_index_start;
1132
1133         /* Outer loop uses 26 flops */
1134     }
1135
1136     /* Increment number of outer iterations */
1137     outeriter        += nri;
1138
1139     /* Update outer/inner flops */
1140
1141     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*413);
1142 }
1143 /*
1144  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW4W4_F_avx_128_fma_double
1145  * Electrostatics interaction: ReactionField
1146  * VdW interaction:            CubicSplineTable
1147  * Geometry:                   Water4-Water4
1148  * Calculate force/pot:        Force
1149  */
1150 void
1151 nb_kernel_ElecRFCut_VdwCSTab_GeomW4W4_F_avx_128_fma_double
1152                     (t_nblist * gmx_restrict                nlist,
1153                      rvec * gmx_restrict                    xx,
1154                      rvec * gmx_restrict                    ff,
1155                      t_forcerec * gmx_restrict              fr,
1156                      t_mdatoms * gmx_restrict               mdatoms,
1157                      nb_kernel_data_t * gmx_restrict        kernel_data,
1158                      t_nrnb * gmx_restrict                  nrnb)
1159 {
1160     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1161      * just 0 for non-waters.
1162      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1163      * jnr indices corresponding to data put in the four positions in the SIMD register.
1164      */
1165     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1166     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1167     int              jnrA,jnrB;
1168     int              j_coord_offsetA,j_coord_offsetB;
1169     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1170     real             rcutoff_scalar;
1171     real             *shiftvec,*fshift,*x,*f;
1172     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1173     int              vdwioffset0;
1174     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1175     int              vdwioffset1;
1176     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1177     int              vdwioffset2;
1178     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1179     int              vdwioffset3;
1180     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1181     int              vdwjidx0A,vdwjidx0B;
1182     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1183     int              vdwjidx1A,vdwjidx1B;
1184     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1185     int              vdwjidx2A,vdwjidx2B;
1186     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1187     int              vdwjidx3A,vdwjidx3B;
1188     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1189     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1190     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1191     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1192     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1193     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1194     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1195     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1196     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1197     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1198     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1199     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1200     real             *charge;
1201     int              nvdwtype;
1202     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1203     int              *vdwtype;
1204     real             *vdwparam;
1205     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1206     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1207     __m128i          vfitab;
1208     __m128i          ifour       = _mm_set1_epi32(4);
1209     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
1210     real             *vftab;
1211     __m128d          dummy_mask,cutoff_mask;
1212     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1213     __m128d          one     = _mm_set1_pd(1.0);
1214     __m128d          two     = _mm_set1_pd(2.0);
1215     x                = xx[0];
1216     f                = ff[0];
1217
1218     nri              = nlist->nri;
1219     iinr             = nlist->iinr;
1220     jindex           = nlist->jindex;
1221     jjnr             = nlist->jjnr;
1222     shiftidx         = nlist->shift;
1223     gid              = nlist->gid;
1224     shiftvec         = fr->shift_vec[0];
1225     fshift           = fr->fshift[0];
1226     facel            = _mm_set1_pd(fr->epsfac);
1227     charge           = mdatoms->chargeA;
1228     krf              = _mm_set1_pd(fr->ic->k_rf);
1229     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1230     crf              = _mm_set1_pd(fr->ic->c_rf);
1231     nvdwtype         = fr->ntype;
1232     vdwparam         = fr->nbfp;
1233     vdwtype          = mdatoms->typeA;
1234
1235     vftab            = kernel_data->table_vdw->data;
1236     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
1237
1238     /* Setup water-specific parameters */
1239     inr              = nlist->iinr[0];
1240     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1241     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1242     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
1243     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1244
1245     jq1              = _mm_set1_pd(charge[inr+1]);
1246     jq2              = _mm_set1_pd(charge[inr+2]);
1247     jq3              = _mm_set1_pd(charge[inr+3]);
1248     vdwjidx0A        = 2*vdwtype[inr+0];
1249     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1250     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1251     qq11             = _mm_mul_pd(iq1,jq1);
1252     qq12             = _mm_mul_pd(iq1,jq2);
1253     qq13             = _mm_mul_pd(iq1,jq3);
1254     qq21             = _mm_mul_pd(iq2,jq1);
1255     qq22             = _mm_mul_pd(iq2,jq2);
1256     qq23             = _mm_mul_pd(iq2,jq3);
1257     qq31             = _mm_mul_pd(iq3,jq1);
1258     qq32             = _mm_mul_pd(iq3,jq2);
1259     qq33             = _mm_mul_pd(iq3,jq3);
1260
1261     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1262     rcutoff_scalar   = fr->rcoulomb;
1263     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1264     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1265
1266     /* Avoid stupid compiler warnings */
1267     jnrA = jnrB = 0;
1268     j_coord_offsetA = 0;
1269     j_coord_offsetB = 0;
1270
1271     outeriter        = 0;
1272     inneriter        = 0;
1273
1274     /* Start outer loop over neighborlists */
1275     for(iidx=0; iidx<nri; iidx++)
1276     {
1277         /* Load shift vector for this list */
1278         i_shift_offset   = DIM*shiftidx[iidx];
1279
1280         /* Load limits for loop over neighbors */
1281         j_index_start    = jindex[iidx];
1282         j_index_end      = jindex[iidx+1];
1283
1284         /* Get outer coordinate index */
1285         inr              = iinr[iidx];
1286         i_coord_offset   = DIM*inr;
1287
1288         /* Load i particle coords and add shift vector */
1289         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1290                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1291
1292         fix0             = _mm_setzero_pd();
1293         fiy0             = _mm_setzero_pd();
1294         fiz0             = _mm_setzero_pd();
1295         fix1             = _mm_setzero_pd();
1296         fiy1             = _mm_setzero_pd();
1297         fiz1             = _mm_setzero_pd();
1298         fix2             = _mm_setzero_pd();
1299         fiy2             = _mm_setzero_pd();
1300         fiz2             = _mm_setzero_pd();
1301         fix3             = _mm_setzero_pd();
1302         fiy3             = _mm_setzero_pd();
1303         fiz3             = _mm_setzero_pd();
1304
1305         /* Start inner kernel loop */
1306         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1307         {
1308
1309             /* Get j neighbor index, and coordinate index */
1310             jnrA             = jjnr[jidx];
1311             jnrB             = jjnr[jidx+1];
1312             j_coord_offsetA  = DIM*jnrA;
1313             j_coord_offsetB  = DIM*jnrB;
1314
1315             /* load j atom coordinates */
1316             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1317                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1318                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1319
1320             /* Calculate displacement vector */
1321             dx00             = _mm_sub_pd(ix0,jx0);
1322             dy00             = _mm_sub_pd(iy0,jy0);
1323             dz00             = _mm_sub_pd(iz0,jz0);
1324             dx11             = _mm_sub_pd(ix1,jx1);
1325             dy11             = _mm_sub_pd(iy1,jy1);
1326             dz11             = _mm_sub_pd(iz1,jz1);
1327             dx12             = _mm_sub_pd(ix1,jx2);
1328             dy12             = _mm_sub_pd(iy1,jy2);
1329             dz12             = _mm_sub_pd(iz1,jz2);
1330             dx13             = _mm_sub_pd(ix1,jx3);
1331             dy13             = _mm_sub_pd(iy1,jy3);
1332             dz13             = _mm_sub_pd(iz1,jz3);
1333             dx21             = _mm_sub_pd(ix2,jx1);
1334             dy21             = _mm_sub_pd(iy2,jy1);
1335             dz21             = _mm_sub_pd(iz2,jz1);
1336             dx22             = _mm_sub_pd(ix2,jx2);
1337             dy22             = _mm_sub_pd(iy2,jy2);
1338             dz22             = _mm_sub_pd(iz2,jz2);
1339             dx23             = _mm_sub_pd(ix2,jx3);
1340             dy23             = _mm_sub_pd(iy2,jy3);
1341             dz23             = _mm_sub_pd(iz2,jz3);
1342             dx31             = _mm_sub_pd(ix3,jx1);
1343             dy31             = _mm_sub_pd(iy3,jy1);
1344             dz31             = _mm_sub_pd(iz3,jz1);
1345             dx32             = _mm_sub_pd(ix3,jx2);
1346             dy32             = _mm_sub_pd(iy3,jy2);
1347             dz32             = _mm_sub_pd(iz3,jz2);
1348             dx33             = _mm_sub_pd(ix3,jx3);
1349             dy33             = _mm_sub_pd(iy3,jy3);
1350             dz33             = _mm_sub_pd(iz3,jz3);
1351
1352             /* Calculate squared distance and things based on it */
1353             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1354             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1355             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1356             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1357             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1358             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1359             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1360             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1361             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1362             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1363
1364             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1365             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1366             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1367             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1368             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1369             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1370             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1371             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1372             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1373             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1374
1375             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1376             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1377             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1378             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1379             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1380             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1381             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1382             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1383             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1384
1385             fjx0             = _mm_setzero_pd();
1386             fjy0             = _mm_setzero_pd();
1387             fjz0             = _mm_setzero_pd();
1388             fjx1             = _mm_setzero_pd();
1389             fjy1             = _mm_setzero_pd();
1390             fjz1             = _mm_setzero_pd();
1391             fjx2             = _mm_setzero_pd();
1392             fjy2             = _mm_setzero_pd();
1393             fjz2             = _mm_setzero_pd();
1394             fjx3             = _mm_setzero_pd();
1395             fjy3             = _mm_setzero_pd();
1396             fjz3             = _mm_setzero_pd();
1397
1398             /**************************
1399              * CALCULATE INTERACTIONS *
1400              **************************/
1401
1402             r00              = _mm_mul_pd(rsq00,rinv00);
1403
1404             /* Calculate table index by multiplying r with table scale and truncate to integer */
1405             rt               = _mm_mul_pd(r00,vftabscale);
1406             vfitab           = _mm_cvttpd_epi32(rt);
1407 #ifdef __XOP__
1408             vfeps            = _mm_frcz_pd(rt);
1409 #else
1410             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
1411 #endif
1412             twovfeps         = _mm_add_pd(vfeps,vfeps);
1413             vfitab           = _mm_slli_epi32(vfitab,3);
1414
1415             /* CUBIC SPLINE TABLE DISPERSION */
1416             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1417             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
1418             GMX_MM_TRANSPOSE2_PD(Y,F);
1419             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1420             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
1421             GMX_MM_TRANSPOSE2_PD(G,H);
1422             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1423             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1424             fvdw6            = _mm_mul_pd(c6_00,FF);
1425
1426             /* CUBIC SPLINE TABLE REPULSION */
1427             vfitab           = _mm_add_epi32(vfitab,ifour);
1428             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1429             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
1430             GMX_MM_TRANSPOSE2_PD(Y,F);
1431             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1432             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
1433             GMX_MM_TRANSPOSE2_PD(G,H);
1434             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1435             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1436             fvdw12           = _mm_mul_pd(c12_00,FF);
1437             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1438
1439             fscal            = fvdw;
1440
1441             /* Update vectorial force */
1442             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1443             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1444             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1445
1446             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1447             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1448             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1449
1450             /**************************
1451              * CALCULATE INTERACTIONS *
1452              **************************/
1453
1454             if (gmx_mm_any_lt(rsq11,rcutoff2))
1455             {
1456
1457             /* REACTION-FIELD ELECTROSTATICS */
1458             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1459
1460             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1461
1462             fscal            = felec;
1463
1464             fscal            = _mm_and_pd(fscal,cutoff_mask);
1465
1466             /* Update vectorial force */
1467             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1468             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1469             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1470
1471             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1472             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1473             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1474
1475             }
1476
1477             /**************************
1478              * CALCULATE INTERACTIONS *
1479              **************************/
1480
1481             if (gmx_mm_any_lt(rsq12,rcutoff2))
1482             {
1483
1484             /* REACTION-FIELD ELECTROSTATICS */
1485             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1486
1487             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1488
1489             fscal            = felec;
1490
1491             fscal            = _mm_and_pd(fscal,cutoff_mask);
1492
1493             /* Update vectorial force */
1494             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1495             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1496             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1497
1498             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1499             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1500             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1501
1502             }
1503
1504             /**************************
1505              * CALCULATE INTERACTIONS *
1506              **************************/
1507
1508             if (gmx_mm_any_lt(rsq13,rcutoff2))
1509             {
1510
1511             /* REACTION-FIELD ELECTROSTATICS */
1512             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1513
1514             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1515
1516             fscal            = felec;
1517
1518             fscal            = _mm_and_pd(fscal,cutoff_mask);
1519
1520             /* Update vectorial force */
1521             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1522             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1523             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1524
1525             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1526             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1527             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1528
1529             }
1530
1531             /**************************
1532              * CALCULATE INTERACTIONS *
1533              **************************/
1534
1535             if (gmx_mm_any_lt(rsq21,rcutoff2))
1536             {
1537
1538             /* REACTION-FIELD ELECTROSTATICS */
1539             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1540
1541             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1542
1543             fscal            = felec;
1544
1545             fscal            = _mm_and_pd(fscal,cutoff_mask);
1546
1547             /* Update vectorial force */
1548             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1549             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1550             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1551
1552             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1553             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1554             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1555
1556             }
1557
1558             /**************************
1559              * CALCULATE INTERACTIONS *
1560              **************************/
1561
1562             if (gmx_mm_any_lt(rsq22,rcutoff2))
1563             {
1564
1565             /* REACTION-FIELD ELECTROSTATICS */
1566             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1567
1568             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1569
1570             fscal            = felec;
1571
1572             fscal            = _mm_and_pd(fscal,cutoff_mask);
1573
1574             /* Update vectorial force */
1575             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1576             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1577             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1578
1579             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1580             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1581             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1582
1583             }
1584
1585             /**************************
1586              * CALCULATE INTERACTIONS *
1587              **************************/
1588
1589             if (gmx_mm_any_lt(rsq23,rcutoff2))
1590             {
1591
1592             /* REACTION-FIELD ELECTROSTATICS */
1593             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1594
1595             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1596
1597             fscal            = felec;
1598
1599             fscal            = _mm_and_pd(fscal,cutoff_mask);
1600
1601             /* Update vectorial force */
1602             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1603             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1604             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1605
1606             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1607             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1608             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1609
1610             }
1611
1612             /**************************
1613              * CALCULATE INTERACTIONS *
1614              **************************/
1615
1616             if (gmx_mm_any_lt(rsq31,rcutoff2))
1617             {
1618
1619             /* REACTION-FIELD ELECTROSTATICS */
1620             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1621
1622             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
1623
1624             fscal            = felec;
1625
1626             fscal            = _mm_and_pd(fscal,cutoff_mask);
1627
1628             /* Update vectorial force */
1629             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1630             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1631             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1632
1633             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1634             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1635             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1636
1637             }
1638
1639             /**************************
1640              * CALCULATE INTERACTIONS *
1641              **************************/
1642
1643             if (gmx_mm_any_lt(rsq32,rcutoff2))
1644             {
1645
1646             /* REACTION-FIELD ELECTROSTATICS */
1647             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1648
1649             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
1650
1651             fscal            = felec;
1652
1653             fscal            = _mm_and_pd(fscal,cutoff_mask);
1654
1655             /* Update vectorial force */
1656             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1657             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1658             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1659
1660             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1661             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1662             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1663
1664             }
1665
1666             /**************************
1667              * CALCULATE INTERACTIONS *
1668              **************************/
1669
1670             if (gmx_mm_any_lt(rsq33,rcutoff2))
1671             {
1672
1673             /* REACTION-FIELD ELECTROSTATICS */
1674             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1675
1676             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
1677
1678             fscal            = felec;
1679
1680             fscal            = _mm_and_pd(fscal,cutoff_mask);
1681
1682             /* Update vectorial force */
1683             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1684             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1685             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1686
1687             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1688             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1689             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1690
1691             }
1692
1693             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);
1694
1695             /* Inner loop uses 351 flops */
1696         }
1697
1698         if(jidx<j_index_end)
1699         {
1700
1701             jnrA             = jjnr[jidx];
1702             j_coord_offsetA  = DIM*jnrA;
1703
1704             /* load j atom coordinates */
1705             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1706                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1707                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1708
1709             /* Calculate displacement vector */
1710             dx00             = _mm_sub_pd(ix0,jx0);
1711             dy00             = _mm_sub_pd(iy0,jy0);
1712             dz00             = _mm_sub_pd(iz0,jz0);
1713             dx11             = _mm_sub_pd(ix1,jx1);
1714             dy11             = _mm_sub_pd(iy1,jy1);
1715             dz11             = _mm_sub_pd(iz1,jz1);
1716             dx12             = _mm_sub_pd(ix1,jx2);
1717             dy12             = _mm_sub_pd(iy1,jy2);
1718             dz12             = _mm_sub_pd(iz1,jz2);
1719             dx13             = _mm_sub_pd(ix1,jx3);
1720             dy13             = _mm_sub_pd(iy1,jy3);
1721             dz13             = _mm_sub_pd(iz1,jz3);
1722             dx21             = _mm_sub_pd(ix2,jx1);
1723             dy21             = _mm_sub_pd(iy2,jy1);
1724             dz21             = _mm_sub_pd(iz2,jz1);
1725             dx22             = _mm_sub_pd(ix2,jx2);
1726             dy22             = _mm_sub_pd(iy2,jy2);
1727             dz22             = _mm_sub_pd(iz2,jz2);
1728             dx23             = _mm_sub_pd(ix2,jx3);
1729             dy23             = _mm_sub_pd(iy2,jy3);
1730             dz23             = _mm_sub_pd(iz2,jz3);
1731             dx31             = _mm_sub_pd(ix3,jx1);
1732             dy31             = _mm_sub_pd(iy3,jy1);
1733             dz31             = _mm_sub_pd(iz3,jz1);
1734             dx32             = _mm_sub_pd(ix3,jx2);
1735             dy32             = _mm_sub_pd(iy3,jy2);
1736             dz32             = _mm_sub_pd(iz3,jz2);
1737             dx33             = _mm_sub_pd(ix3,jx3);
1738             dy33             = _mm_sub_pd(iy3,jy3);
1739             dz33             = _mm_sub_pd(iz3,jz3);
1740
1741             /* Calculate squared distance and things based on it */
1742             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1743             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1744             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1745             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1746             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1747             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1748             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1749             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1750             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1751             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1752
1753             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1754             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1755             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1756             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1757             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1758             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1759             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1760             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1761             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1762             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1763
1764             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1765             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1766             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1767             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1768             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1769             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1770             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1771             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1772             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1773
1774             fjx0             = _mm_setzero_pd();
1775             fjy0             = _mm_setzero_pd();
1776             fjz0             = _mm_setzero_pd();
1777             fjx1             = _mm_setzero_pd();
1778             fjy1             = _mm_setzero_pd();
1779             fjz1             = _mm_setzero_pd();
1780             fjx2             = _mm_setzero_pd();
1781             fjy2             = _mm_setzero_pd();
1782             fjz2             = _mm_setzero_pd();
1783             fjx3             = _mm_setzero_pd();
1784             fjy3             = _mm_setzero_pd();
1785             fjz3             = _mm_setzero_pd();
1786
1787             /**************************
1788              * CALCULATE INTERACTIONS *
1789              **************************/
1790
1791             r00              = _mm_mul_pd(rsq00,rinv00);
1792
1793             /* Calculate table index by multiplying r with table scale and truncate to integer */
1794             rt               = _mm_mul_pd(r00,vftabscale);
1795             vfitab           = _mm_cvttpd_epi32(rt);
1796 #ifdef __XOP__
1797             vfeps            = _mm_frcz_pd(rt);
1798 #else
1799             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
1800 #endif
1801             twovfeps         = _mm_add_pd(vfeps,vfeps);
1802             vfitab           = _mm_slli_epi32(vfitab,3);
1803
1804             /* CUBIC SPLINE TABLE DISPERSION */
1805             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1806             F                = _mm_setzero_pd();
1807             GMX_MM_TRANSPOSE2_PD(Y,F);
1808             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1809             H                = _mm_setzero_pd();
1810             GMX_MM_TRANSPOSE2_PD(G,H);
1811             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1812             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1813             fvdw6            = _mm_mul_pd(c6_00,FF);
1814
1815             /* CUBIC SPLINE TABLE REPULSION */
1816             vfitab           = _mm_add_epi32(vfitab,ifour);
1817             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1818             F                = _mm_setzero_pd();
1819             GMX_MM_TRANSPOSE2_PD(Y,F);
1820             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1821             H                = _mm_setzero_pd();
1822             GMX_MM_TRANSPOSE2_PD(G,H);
1823             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1824             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1825             fvdw12           = _mm_mul_pd(c12_00,FF);
1826             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1827
1828             fscal            = fvdw;
1829
1830             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1831
1832             /* Update vectorial force */
1833             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1834             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1835             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1836
1837             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1838             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1839             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1840
1841             /**************************
1842              * CALCULATE INTERACTIONS *
1843              **************************/
1844
1845             if (gmx_mm_any_lt(rsq11,rcutoff2))
1846             {
1847
1848             /* REACTION-FIELD ELECTROSTATICS */
1849             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1850
1851             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1852
1853             fscal            = felec;
1854
1855             fscal            = _mm_and_pd(fscal,cutoff_mask);
1856
1857             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1858
1859             /* Update vectorial force */
1860             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1861             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1862             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1863
1864             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1865             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1866             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1867
1868             }
1869
1870             /**************************
1871              * CALCULATE INTERACTIONS *
1872              **************************/
1873
1874             if (gmx_mm_any_lt(rsq12,rcutoff2))
1875             {
1876
1877             /* REACTION-FIELD ELECTROSTATICS */
1878             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1879
1880             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1881
1882             fscal            = felec;
1883
1884             fscal            = _mm_and_pd(fscal,cutoff_mask);
1885
1886             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1887
1888             /* Update vectorial force */
1889             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1890             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1891             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1892
1893             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1894             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1895             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1896
1897             }
1898
1899             /**************************
1900              * CALCULATE INTERACTIONS *
1901              **************************/
1902
1903             if (gmx_mm_any_lt(rsq13,rcutoff2))
1904             {
1905
1906             /* REACTION-FIELD ELECTROSTATICS */
1907             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1908
1909             cutoff_mask      = _mm_cmplt_pd(rsq13,rcutoff2);
1910
1911             fscal            = felec;
1912
1913             fscal            = _mm_and_pd(fscal,cutoff_mask);
1914
1915             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1916
1917             /* Update vectorial force */
1918             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1919             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1920             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1921
1922             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1923             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1924             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1925
1926             }
1927
1928             /**************************
1929              * CALCULATE INTERACTIONS *
1930              **************************/
1931
1932             if (gmx_mm_any_lt(rsq21,rcutoff2))
1933             {
1934
1935             /* REACTION-FIELD ELECTROSTATICS */
1936             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1937
1938             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1939
1940             fscal            = felec;
1941
1942             fscal            = _mm_and_pd(fscal,cutoff_mask);
1943
1944             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1945
1946             /* Update vectorial force */
1947             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1948             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1949             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1950
1951             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1952             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1953             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1954
1955             }
1956
1957             /**************************
1958              * CALCULATE INTERACTIONS *
1959              **************************/
1960
1961             if (gmx_mm_any_lt(rsq22,rcutoff2))
1962             {
1963
1964             /* REACTION-FIELD ELECTROSTATICS */
1965             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1966
1967             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1968
1969             fscal            = felec;
1970
1971             fscal            = _mm_and_pd(fscal,cutoff_mask);
1972
1973             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1974
1975             /* Update vectorial force */
1976             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1977             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1978             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1979
1980             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1981             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1982             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1983
1984             }
1985
1986             /**************************
1987              * CALCULATE INTERACTIONS *
1988              **************************/
1989
1990             if (gmx_mm_any_lt(rsq23,rcutoff2))
1991             {
1992
1993             /* REACTION-FIELD ELECTROSTATICS */
1994             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1995
1996             cutoff_mask      = _mm_cmplt_pd(rsq23,rcutoff2);
1997
1998             fscal            = felec;
1999
2000             fscal            = _mm_and_pd(fscal,cutoff_mask);
2001
2002             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2003
2004             /* Update vectorial force */
2005             fix2             = _mm_macc_pd(dx23,fscal,fix2);
2006             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
2007             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
2008
2009             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
2010             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
2011             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
2012
2013             }
2014
2015             /**************************
2016              * CALCULATE INTERACTIONS *
2017              **************************/
2018
2019             if (gmx_mm_any_lt(rsq31,rcutoff2))
2020             {
2021
2022             /* REACTION-FIELD ELECTROSTATICS */
2023             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
2024
2025             cutoff_mask      = _mm_cmplt_pd(rsq31,rcutoff2);
2026
2027             fscal            = felec;
2028
2029             fscal            = _mm_and_pd(fscal,cutoff_mask);
2030
2031             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2032
2033             /* Update vectorial force */
2034             fix3             = _mm_macc_pd(dx31,fscal,fix3);
2035             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
2036             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
2037
2038             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
2039             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
2040             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
2041
2042             }
2043
2044             /**************************
2045              * CALCULATE INTERACTIONS *
2046              **************************/
2047
2048             if (gmx_mm_any_lt(rsq32,rcutoff2))
2049             {
2050
2051             /* REACTION-FIELD ELECTROSTATICS */
2052             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
2053
2054             cutoff_mask      = _mm_cmplt_pd(rsq32,rcutoff2);
2055
2056             fscal            = felec;
2057
2058             fscal            = _mm_and_pd(fscal,cutoff_mask);
2059
2060             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2061
2062             /* Update vectorial force */
2063             fix3             = _mm_macc_pd(dx32,fscal,fix3);
2064             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
2065             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
2066
2067             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
2068             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
2069             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
2070
2071             }
2072
2073             /**************************
2074              * CALCULATE INTERACTIONS *
2075              **************************/
2076
2077             if (gmx_mm_any_lt(rsq33,rcutoff2))
2078             {
2079
2080             /* REACTION-FIELD ELECTROSTATICS */
2081             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
2082
2083             cutoff_mask      = _mm_cmplt_pd(rsq33,rcutoff2);
2084
2085             fscal            = felec;
2086
2087             fscal            = _mm_and_pd(fscal,cutoff_mask);
2088
2089             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
2090
2091             /* Update vectorial force */
2092             fix3             = _mm_macc_pd(dx33,fscal,fix3);
2093             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
2094             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
2095
2096             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
2097             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
2098             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
2099
2100             }
2101
2102             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2103
2104             /* Inner loop uses 351 flops */
2105         }
2106
2107         /* End of innermost loop */
2108
2109         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2110                                               f+i_coord_offset,fshift+i_shift_offset);
2111
2112         /* Increment number of inner iterations */
2113         inneriter                  += j_index_end - j_index_start;
2114
2115         /* Outer loop uses 24 flops */
2116     }
2117
2118     /* Increment number of outer iterations */
2119     outeriter        += nri;
2120
2121     /* Update outer/inner flops */
2122
2123     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*351);
2124 }