src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_avx_128_fma_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_double kernel generator.
  37  */
  38 #include "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_double.h"
  48 #include "kernelutil_x86_avx_128_fma_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_VF_avx_128_fma_double
  52  * Electrostatics interaction: ReactionField
  53  * VdW interaction:            CubicSplineTable
  54  * Geometry:                   Water3-Water3
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_VF_avx_128_fma_double
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB;
  75     int              j_coord_offsetA,j_coord_offsetB;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  80     int              vdwioffset0;
  81     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  82     int              vdwioffset1;
  83     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  84     int              vdwioffset2;
  85     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  86     int              vdwjidx0A,vdwjidx0B;
  87     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  88     int              vdwjidx1A,vdwjidx1B;
  89     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  90     int              vdwjidx2A,vdwjidx2B;
  91     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  92     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  93     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  94     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  95     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  96     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  97     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  98     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  99     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 100     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 101     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 102     real             *charge;
 103     int              nvdwtype;
 104     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 105     int              *vdwtype;
 106     real             *vdwparam;
 107     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 108     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 109     __m128i          vfitab;
 110     __m128i          ifour       = _mm_set1_epi32(4);
 111     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
 112     real             *vftab;
 113     __m128d          dummy_mask,cutoff_mask;
 114     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 115     __m128d          one     = _mm_set1_pd(1.0);
 116     __m128d          two     = _mm_set1_pd(2.0);
 117     x                = xx[0];
 118     f                = ff[0];
 119
 120     nri              = nlist->nri;
 121     iinr             = nlist->iinr;
 122     jindex           = nlist->jindex;
 123     jjnr             = nlist->jjnr;
 124     shiftidx         = nlist->shift;
 125     gid              = nlist->gid;
 126     shiftvec         = fr->shift_vec[0];
 127     fshift           = fr->fshift[0];
 128     facel            = _mm_set1_pd(fr->epsfac);
 129     charge           = mdatoms->chargeA;
 130     krf              = _mm_set1_pd(fr->ic->k_rf);
 131     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 132     crf              = _mm_set1_pd(fr->ic->c_rf);
 133     nvdwtype         = fr->ntype;
 134     vdwparam         = fr->nbfp;
 135     vdwtype          = mdatoms->typeA;
 136
 137     vftab            = kernel_data->table_vdw->data;
 138     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
 139
 140     /* Setup water-specific parameters */
 141     inr              = nlist->iinr[0];
 142     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 143     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 144     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 145     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 146
 147     jq0              = _mm_set1_pd(charge[inr+0]);
 148     jq1              = _mm_set1_pd(charge[inr+1]);
 149     jq2              = _mm_set1_pd(charge[inr+2]);
 150     vdwjidx0A        = 2*vdwtype[inr+0];
 151     qq00             = _mm_mul_pd(iq0,jq0);
 152     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 153     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 154     qq01             = _mm_mul_pd(iq0,jq1);
 155     qq02             = _mm_mul_pd(iq0,jq2);
 156     qq10             = _mm_mul_pd(iq1,jq0);
 157     qq11             = _mm_mul_pd(iq1,jq1);
 158     qq12             = _mm_mul_pd(iq1,jq2);
 159     qq20             = _mm_mul_pd(iq2,jq0);
 160     qq21             = _mm_mul_pd(iq2,jq1);
 161     qq22             = _mm_mul_pd(iq2,jq2);
 162
 163     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 164     rcutoff_scalar   = fr->rcoulomb;
 165     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 166     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 167
 168     /* Avoid stupid compiler warnings */
 169     jnrA = jnrB = 0;
 170     j_coord_offsetA = 0;
 171     j_coord_offsetB = 0;
 172
 173     outeriter        = 0;
 174     inneriter        = 0;
 175
 176     /* Start outer loop over neighborlists */
 177     for(iidx=0; iidx<nri; iidx++)
 178     {
 179         /* Load shift vector for this list */
 180         i_shift_offset   = DIM*shiftidx[iidx];
 181
 182         /* Load limits for loop over neighbors */
 183         j_index_start    = jindex[iidx];
 184         j_index_end      = jindex[iidx+1];
 185
 186         /* Get outer coordinate index */
 187         inr              = iinr[iidx];
 188         i_coord_offset   = DIM*inr;
 189
 190         /* Load i particle coords and add shift vector */
 191         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 192                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 193
 194         fix0             = _mm_setzero_pd();
 195         fiy0             = _mm_setzero_pd();
 196         fiz0             = _mm_setzero_pd();
 197         fix1             = _mm_setzero_pd();
 198         fiy1             = _mm_setzero_pd();
 199         fiz1             = _mm_setzero_pd();
 200         fix2             = _mm_setzero_pd();
 201         fiy2             = _mm_setzero_pd();
 202         fiz2             = _mm_setzero_pd();
 203
 204         /* Reset potential sums */
 205         velecsum         = _mm_setzero_pd();
 206         vvdwsum          = _mm_setzero_pd();
 207
 208         /* Start inner kernel loop */
 209         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 210         {
 211
 212             /* Get j neighbor index, and coordinate index */
 213             jnrA             = jjnr[jidx];
 214             jnrB             = jjnr[jidx+1];
 215             j_coord_offsetA  = DIM*jnrA;
 216             j_coord_offsetB  = DIM*jnrB;
 217
 218             /* load j atom coordinates */
 219             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 220                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 221
 222             /* Calculate displacement vector */
 223             dx00             = _mm_sub_pd(ix0,jx0);
 224             dy00             = _mm_sub_pd(iy0,jy0);
 225             dz00             = _mm_sub_pd(iz0,jz0);
 226             dx01             = _mm_sub_pd(ix0,jx1);
 227             dy01             = _mm_sub_pd(iy0,jy1);
 228             dz01             = _mm_sub_pd(iz0,jz1);
 229             dx02             = _mm_sub_pd(ix0,jx2);
 230             dy02             = _mm_sub_pd(iy0,jy2);
 231             dz02             = _mm_sub_pd(iz0,jz2);
 232             dx10             = _mm_sub_pd(ix1,jx0);
 233             dy10             = _mm_sub_pd(iy1,jy0);
 234             dz10             = _mm_sub_pd(iz1,jz0);
 235             dx11             = _mm_sub_pd(ix1,jx1);
 236             dy11             = _mm_sub_pd(iy1,jy1);
 237             dz11             = _mm_sub_pd(iz1,jz1);
 238             dx12             = _mm_sub_pd(ix1,jx2);
 239             dy12             = _mm_sub_pd(iy1,jy2);
 240             dz12             = _mm_sub_pd(iz1,jz2);
 241             dx20             = _mm_sub_pd(ix2,jx0);
 242             dy20             = _mm_sub_pd(iy2,jy0);
 243             dz20             = _mm_sub_pd(iz2,jz0);
 244             dx21             = _mm_sub_pd(ix2,jx1);
 245             dy21             = _mm_sub_pd(iy2,jy1);
 246             dz21             = _mm_sub_pd(iz2,jz1);
 247             dx22             = _mm_sub_pd(ix2,jx2);
 248             dy22             = _mm_sub_pd(iy2,jy2);
 249             dz22             = _mm_sub_pd(iz2,jz2);
 250
 251             /* Calculate squared distance and things based on it */
 252             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 253             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 254             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 255             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 256             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 257             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 258             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 259             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 260             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 261
 262             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 263             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 264             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 265             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 266             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 267             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 268             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 269             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 270             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 271
 272             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 273             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 274             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 275             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 276             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 277             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 278             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 279             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 280             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 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
 292             /**************************
 293              * CALCULATE INTERACTIONS *
 294              **************************/
 295
 296             if (gmx_mm_any_lt(rsq00,rcutoff2))
 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             /* REACTION-FIELD ELECTROSTATICS */
 313             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
 314             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 315
 316             /* CUBIC SPLINE TABLE DISPERSION */
 317             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 318             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 319             GMX_MM_TRANSPOSE2_PD(Y,F);
 320             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 321             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
 322             GMX_MM_TRANSPOSE2_PD(G,H);
 323             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 324             VV               = _mm_macc_pd(vfeps,Fp,Y);
 325             vvdw6            = _mm_mul_pd(c6_00,VV);
 326             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 327             fvdw6            = _mm_mul_pd(c6_00,FF);
 328
 329             /* CUBIC SPLINE TABLE REPULSION */
 330             vfitab           = _mm_add_epi32(vfitab,ifour);
 331             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 332             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
 333             GMX_MM_TRANSPOSE2_PD(Y,F);
 334             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 335             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
 336             GMX_MM_TRANSPOSE2_PD(G,H);
 337             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 338             VV               = _mm_macc_pd(vfeps,Fp,Y);
 339             vvdw12           = _mm_mul_pd(c12_00,VV);
 340             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 341             fvdw12           = _mm_mul_pd(c12_00,FF);
 342             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 343             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 344
 345             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 346
 347             /* Update potential sum for this i atom from the interaction with this j atom. */
 348             velec            = _mm_and_pd(velec,cutoff_mask);
 349             velecsum         = _mm_add_pd(velecsum,velec);
 350             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 351             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 352
 353             fscal            = _mm_add_pd(felec,fvdw);
 354
 355             fscal            = _mm_and_pd(fscal,cutoff_mask);
 356
 357             /* Update vectorial force */
 358             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 359             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 360             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 361
 362             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 363             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 364             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 365
 366             }
 367
 368             /**************************
 369              * CALCULATE INTERACTIONS *
 370              **************************/
 371
 372             if (gmx_mm_any_lt(rsq01,rcutoff2))
 373             {
 374
 375             /* REACTION-FIELD ELECTROSTATICS */
 376             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_macc_pd(krf,rsq01,rinv01),crf));
 377             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
 378
 379             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
 380
 381             /* Update potential sum for this i atom from the interaction with this j atom. */
 382             velec            = _mm_and_pd(velec,cutoff_mask);
 383             velecsum         = _mm_add_pd(velecsum,velec);
 384
 385             fscal            = felec;
 386
 387             fscal            = _mm_and_pd(fscal,cutoff_mask);
 388
 389             /* Update vectorial force */
 390             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 391             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 392             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 393
 394             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 395             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 396             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 397
 398             }
 399
 400             /**************************
 401              * CALCULATE INTERACTIONS *
 402              **************************/
 403
 404             if (gmx_mm_any_lt(rsq02,rcutoff2))
 405             {
 406
 407             /* REACTION-FIELD ELECTROSTATICS */
 408             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_macc_pd(krf,rsq02,rinv02),crf));
 409             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
 410
 411             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
 412
 413             /* Update potential sum for this i atom from the interaction with this j atom. */
 414             velec            = _mm_and_pd(velec,cutoff_mask);
 415             velecsum         = _mm_add_pd(velecsum,velec);
 416
 417             fscal            = felec;
 418
 419             fscal            = _mm_and_pd(fscal,cutoff_mask);
 420
 421             /* Update vectorial force */
 422             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 423             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 424             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 425
 426             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 427             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 428             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 429
 430             }
 431
 432             /**************************
 433              * CALCULATE INTERACTIONS *
 434              **************************/
 435
 436             if (gmx_mm_any_lt(rsq10,rcutoff2))
 437             {
 438
 439             /* REACTION-FIELD ELECTROSTATICS */
 440             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
 441             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 442
 443             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
 444
 445             /* Update potential sum for this i atom from the interaction with this j atom. */
 446             velec            = _mm_and_pd(velec,cutoff_mask);
 447             velecsum         = _mm_add_pd(velecsum,velec);
 448
 449             fscal            = felec;
 450
 451             fscal            = _mm_and_pd(fscal,cutoff_mask);
 452
 453             /* Update vectorial force */
 454             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 455             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 456             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 457
 458             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 459             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 460             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 461
 462             }
 463
 464             /**************************
 465              * CALCULATE INTERACTIONS *
 466              **************************/
 467
 468             if (gmx_mm_any_lt(rsq11,rcutoff2))
 469             {
 470
 471             /* REACTION-FIELD ELECTROSTATICS */
 472             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 473             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 474
 475             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 476
 477             /* Update potential sum for this i atom from the interaction with this j atom. */
 478             velec            = _mm_and_pd(velec,cutoff_mask);
 479             velecsum         = _mm_add_pd(velecsum,velec);
 480
 481             fscal            = felec;
 482
 483             fscal            = _mm_and_pd(fscal,cutoff_mask);
 484
 485             /* Update vectorial force */
 486             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 487             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 488             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 489
 490             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 491             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 492             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 493
 494             }
 495
 496             /**************************
 497              * CALCULATE INTERACTIONS *
 498              **************************/
 499
 500             if (gmx_mm_any_lt(rsq12,rcutoff2))
 501             {
 502
 503             /* REACTION-FIELD ELECTROSTATICS */
 504             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 505             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 506
 507             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 508
 509             /* Update potential sum for this i atom from the interaction with this j atom. */
 510             velec            = _mm_and_pd(velec,cutoff_mask);
 511             velecsum         = _mm_add_pd(velecsum,velec);
 512
 513             fscal            = felec;
 514
 515             fscal            = _mm_and_pd(fscal,cutoff_mask);
 516
 517             /* Update vectorial force */
 518             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 519             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 520             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 521
 522             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 523             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 524             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 525
 526             }
 527
 528             /**************************
 529              * CALCULATE INTERACTIONS *
 530              **************************/
 531
 532             if (gmx_mm_any_lt(rsq20,rcutoff2))
 533             {
 534
 535             /* REACTION-FIELD ELECTROSTATICS */
 536             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
 537             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 538
 539             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
 540
 541             /* Update potential sum for this i atom from the interaction with this j atom. */
 542             velec            = _mm_and_pd(velec,cutoff_mask);
 543             velecsum         = _mm_add_pd(velecsum,velec);
 544
 545             fscal            = felec;
 546
 547             fscal            = _mm_and_pd(fscal,cutoff_mask);
 548
 549             /* Update vectorial force */
 550             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 551             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 552             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 553
 554             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 555             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 556             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 557
 558             }
 559
 560             /**************************
 561              * CALCULATE INTERACTIONS *
 562              **************************/
 563
 564             if (gmx_mm_any_lt(rsq21,rcutoff2))
 565             {
 566
 567             /* REACTION-FIELD ELECTROSTATICS */
 568             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 569             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 570
 571             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 572
 573             /* Update potential sum for this i atom from the interaction with this j atom. */
 574             velec            = _mm_and_pd(velec,cutoff_mask);
 575             velecsum         = _mm_add_pd(velecsum,velec);
 576
 577             fscal            = felec;
 578
 579             fscal            = _mm_and_pd(fscal,cutoff_mask);
 580
 581             /* Update vectorial force */
 582             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 583             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 584             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 585
 586             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 587             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 588             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 589
 590             }
 591
 592             /**************************
 593              * CALCULATE INTERACTIONS *
 594              **************************/
 595
 596             if (gmx_mm_any_lt(rsq22,rcutoff2))
 597             {
 598
 599             /* REACTION-FIELD ELECTROSTATICS */
 600             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 601             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 602
 603             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 604
 605             /* Update potential sum for this i atom from the interaction with this j atom. */
 606             velec            = _mm_and_pd(velec,cutoff_mask);
 607             velecsum         = _mm_add_pd(velecsum,velec);
 608
 609             fscal            = felec;
 610
 611             fscal            = _mm_and_pd(fscal,cutoff_mask);
 612
 613             /* Update vectorial force */
 614             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 615             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 616             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 617
 618             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 619             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 620             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 621
 622             }
 623
 624             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 625
 626             /* Inner loop uses 387 flops */
 627         }
 628
 629         if(jidx<j_index_end)
 630         {
 631
 632             jnrA             = jjnr[jidx];
 633             j_coord_offsetA  = DIM*jnrA;
 634
 635             /* load j atom coordinates */
 636             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 637                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 638
 639             /* Calculate displacement vector */
 640             dx00             = _mm_sub_pd(ix0,jx0);
 641             dy00             = _mm_sub_pd(iy0,jy0);
 642             dz00             = _mm_sub_pd(iz0,jz0);
 643             dx01             = _mm_sub_pd(ix0,jx1);
 644             dy01             = _mm_sub_pd(iy0,jy1);
 645             dz01             = _mm_sub_pd(iz0,jz1);
 646             dx02             = _mm_sub_pd(ix0,jx2);
 647             dy02             = _mm_sub_pd(iy0,jy2);
 648             dz02             = _mm_sub_pd(iz0,jz2);
 649             dx10             = _mm_sub_pd(ix1,jx0);
 650             dy10             = _mm_sub_pd(iy1,jy0);
 651             dz10             = _mm_sub_pd(iz1,jz0);
 652             dx11             = _mm_sub_pd(ix1,jx1);
 653             dy11             = _mm_sub_pd(iy1,jy1);
 654             dz11             = _mm_sub_pd(iz1,jz1);
 655             dx12             = _mm_sub_pd(ix1,jx2);
 656             dy12             = _mm_sub_pd(iy1,jy2);
 657             dz12             = _mm_sub_pd(iz1,jz2);
 658             dx20             = _mm_sub_pd(ix2,jx0);
 659             dy20             = _mm_sub_pd(iy2,jy0);
 660             dz20             = _mm_sub_pd(iz2,jz0);
 661             dx21             = _mm_sub_pd(ix2,jx1);
 662             dy21             = _mm_sub_pd(iy2,jy1);
 663             dz21             = _mm_sub_pd(iz2,jz1);
 664             dx22             = _mm_sub_pd(ix2,jx2);
 665             dy22             = _mm_sub_pd(iy2,jy2);
 666             dz22             = _mm_sub_pd(iz2,jz2);
 667
 668             /* Calculate squared distance and things based on it */
 669             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 670             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 671             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 672             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 673             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 674             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 675             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 676             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 677             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 678
 679             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 680             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 681             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 682             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 683             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 684             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 685             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 686             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 687             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 688
 689             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 690             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 691             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 692             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 693             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 694             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 695             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 696             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 697             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 698
 699             fjx0             = _mm_setzero_pd();
 700             fjy0             = _mm_setzero_pd();
 701             fjz0             = _mm_setzero_pd();
 702             fjx1             = _mm_setzero_pd();
 703             fjy1             = _mm_setzero_pd();
 704             fjz1             = _mm_setzero_pd();
 705             fjx2             = _mm_setzero_pd();
 706             fjy2             = _mm_setzero_pd();
 707             fjz2             = _mm_setzero_pd();
 708
 709             /**************************
 710              * CALCULATE INTERACTIONS *
 711              **************************/
 712
 713             if (gmx_mm_any_lt(rsq00,rcutoff2))
 714             {
 715
 716             r00              = _mm_mul_pd(rsq00,rinv00);
 717
 718             /* Calculate table index by multiplying r with table scale and truncate to integer */
 719             rt               = _mm_mul_pd(r00,vftabscale);
 720             vfitab           = _mm_cvttpd_epi32(rt);
 721 #ifdef __XOP__
 722             vfeps            = _mm_frcz_pd(rt);
 723 #else
 724             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
 725 #endif
 726             twovfeps         = _mm_add_pd(vfeps,vfeps);
 727             vfitab           = _mm_slli_epi32(vfitab,3);
 728
 729             /* REACTION-FIELD ELECTROSTATICS */
 730             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
 731             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 732
 733             /* CUBIC SPLINE TABLE DISPERSION */
 734             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 735             F                = _mm_setzero_pd();
 736             GMX_MM_TRANSPOSE2_PD(Y,F);
 737             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 738             H                = _mm_setzero_pd();
 739             GMX_MM_TRANSPOSE2_PD(G,H);
 740             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 741             VV               = _mm_macc_pd(vfeps,Fp,Y);
 742             vvdw6            = _mm_mul_pd(c6_00,VV);
 743             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 744             fvdw6            = _mm_mul_pd(c6_00,FF);
 745
 746             /* CUBIC SPLINE TABLE REPULSION */
 747             vfitab           = _mm_add_epi32(vfitab,ifour);
 748             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
 749             F                = _mm_setzero_pd();
 750             GMX_MM_TRANSPOSE2_PD(Y,F);
 751             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
 752             H                = _mm_setzero_pd();
 753             GMX_MM_TRANSPOSE2_PD(G,H);
 754             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
 755             VV               = _mm_macc_pd(vfeps,Fp,Y);
 756             vvdw12           = _mm_mul_pd(c12_00,VV);
 757             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
 758             fvdw12           = _mm_mul_pd(c12_00,FF);
 759             vvdw             = _mm_add_pd(vvdw12,vvdw6);
 760             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
 761
 762             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 763
 764             /* Update potential sum for this i atom from the interaction with this j atom. */
 765             velec            = _mm_and_pd(velec,cutoff_mask);
 766             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 767             velecsum         = _mm_add_pd(velecsum,velec);
 768             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 769             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 770             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 771
 772             fscal            = _mm_add_pd(felec,fvdw);
 773
 774             fscal            = _mm_and_pd(fscal,cutoff_mask);
 775
 776             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 777
 778             /* Update vectorial force */
 779             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 780             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 781             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 782
 783             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 784             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 785             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 786
 787             }
 788
 789             /**************************
 790              * CALCULATE INTERACTIONS *
 791              **************************/
 792
 793             if (gmx_mm_any_lt(rsq01,rcutoff2))
 794             {
 795
 796             /* REACTION-FIELD ELECTROSTATICS */
 797             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_macc_pd(krf,rsq01,rinv01),crf));
 798             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
 799
 800             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
 801
 802             /* Update potential sum for this i atom from the interaction with this j atom. */
 803             velec            = _mm_and_pd(velec,cutoff_mask);
 804             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 805             velecsum         = _mm_add_pd(velecsum,velec);
 806
 807             fscal            = felec;
 808
 809             fscal            = _mm_and_pd(fscal,cutoff_mask);
 810
 811             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 812
 813             /* Update vectorial force */
 814             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 815             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 816             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 817
 818             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 819             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 820             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 821
 822             }
 823
 824             /**************************
 825              * CALCULATE INTERACTIONS *
 826              **************************/
 827
 828             if (gmx_mm_any_lt(rsq02,rcutoff2))
 829             {
 830
 831             /* REACTION-FIELD ELECTROSTATICS */
 832             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_macc_pd(krf,rsq02,rinv02),crf));
 833             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
 834
 835             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
 836
 837             /* Update potential sum for this i atom from the interaction with this j atom. */
 838             velec            = _mm_and_pd(velec,cutoff_mask);
 839             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 840             velecsum         = _mm_add_pd(velecsum,velec);
 841
 842             fscal            = felec;
 843
 844             fscal            = _mm_and_pd(fscal,cutoff_mask);
 845
 846             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 847
 848             /* Update vectorial force */
 849             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 850             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 851             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 852
 853             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 854             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 855             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 856
 857             }
 858
 859             /**************************
 860              * CALCULATE INTERACTIONS *
 861              **************************/
 862
 863             if (gmx_mm_any_lt(rsq10,rcutoff2))
 864             {
 865
 866             /* REACTION-FIELD ELECTROSTATICS */
 867             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
 868             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 869
 870             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
 871
 872             /* Update potential sum for this i atom from the interaction with this j atom. */
 873             velec            = _mm_and_pd(velec,cutoff_mask);
 874             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 875             velecsum         = _mm_add_pd(velecsum,velec);
 876
 877             fscal            = felec;
 878
 879             fscal            = _mm_and_pd(fscal,cutoff_mask);
 880
 881             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 882
 883             /* Update vectorial force */
 884             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 885             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 886             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 887
 888             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 889             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 890             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 891
 892             }
 893
 894             /**************************
 895              * CALCULATE INTERACTIONS *
 896              **************************/
 897
 898             if (gmx_mm_any_lt(rsq11,rcutoff2))
 899             {
 900
 901             /* REACTION-FIELD ELECTROSTATICS */
 902             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 903             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 904
 905             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 906
 907             /* Update potential sum for this i atom from the interaction with this j atom. */
 908             velec            = _mm_and_pd(velec,cutoff_mask);
 909             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 910             velecsum         = _mm_add_pd(velecsum,velec);
 911
 912             fscal            = felec;
 913
 914             fscal            = _mm_and_pd(fscal,cutoff_mask);
 915
 916             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 917
 918             /* Update vectorial force */
 919             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 920             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 921             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 922
 923             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 924             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 925             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 926
 927             }
 928
 929             /**************************
 930              * CALCULATE INTERACTIONS *
 931              **************************/
 932
 933             if (gmx_mm_any_lt(rsq12,rcutoff2))
 934             {
 935
 936             /* REACTION-FIELD ELECTROSTATICS */
 937             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 938             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 939
 940             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 941
 942             /* Update potential sum for this i atom from the interaction with this j atom. */
 943             velec            = _mm_and_pd(velec,cutoff_mask);
 944             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 945             velecsum         = _mm_add_pd(velecsum,velec);
 946
 947             fscal            = felec;
 948
 949             fscal            = _mm_and_pd(fscal,cutoff_mask);
 950
 951             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 952
 953             /* Update vectorial force */
 954             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 955             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 956             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 957
 958             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 959             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 960             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 961
 962             }
 963
 964             /**************************
 965              * CALCULATE INTERACTIONS *
 966              **************************/
 967
 968             if (gmx_mm_any_lt(rsq20,rcutoff2))
 969             {
 970
 971             /* REACTION-FIELD ELECTROSTATICS */
 972             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
 973             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 974
 975             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
 976
 977             /* Update potential sum for this i atom from the interaction with this j atom. */
 978             velec            = _mm_and_pd(velec,cutoff_mask);
 979             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 980             velecsum         = _mm_add_pd(velecsum,velec);
 981
 982             fscal            = felec;
 983
 984             fscal            = _mm_and_pd(fscal,cutoff_mask);
 985
 986             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 987
 988             /* Update vectorial force */
 989             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 990             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 991             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 992
 993             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 994             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 995             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 996
 997             }
 998
 999             /**************************
1000              * CALCULATE INTERACTIONS *
1001              **************************/
1002
1003             if (gmx_mm_any_lt(rsq21,rcutoff2))
1004             {
1005
1006             /* REACTION-FIELD ELECTROSTATICS */
1007             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
1008             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1009
1010             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1011
1012             /* Update potential sum for this i atom from the interaction with this j atom. */
1013             velec            = _mm_and_pd(velec,cutoff_mask);
1014             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1015             velecsum         = _mm_add_pd(velecsum,velec);
1016
1017             fscal            = felec;
1018
1019             fscal            = _mm_and_pd(fscal,cutoff_mask);
1020
1021             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1022
1023             /* Update vectorial force */
1024             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1025             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1026             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1027
1028             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1029             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1030             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1031
1032             }
1033
1034             /**************************
1035              * CALCULATE INTERACTIONS *
1036              **************************/
1037
1038             if (gmx_mm_any_lt(rsq22,rcutoff2))
1039             {
1040
1041             /* REACTION-FIELD ELECTROSTATICS */
1042             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
1043             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1044
1045             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1046
1047             /* Update potential sum for this i atom from the interaction with this j atom. */
1048             velec            = _mm_and_pd(velec,cutoff_mask);
1049             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
1050             velecsum         = _mm_add_pd(velecsum,velec);
1051
1052             fscal            = felec;
1053
1054             fscal            = _mm_and_pd(fscal,cutoff_mask);
1055
1056             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1057
1058             /* Update vectorial force */
1059             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1060             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1061             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1062
1063             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1064             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1065             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1066
1067             }
1068
1069             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1070
1071             /* Inner loop uses 387 flops */
1072         }
1073
1074         /* End of innermost loop */
1075
1076         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1077                                               f+i_coord_offset,fshift+i_shift_offset);
1078
1079         ggid                        = gid[iidx];
1080         /* Update potential energies */
1081         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1082         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
1083
1084         /* Increment number of inner iterations */
1085         inneriter                  += j_index_end - j_index_start;
1086
1087         /* Outer loop uses 20 flops */
1088     }
1089
1090     /* Increment number of outer iterations */
1091     outeriter        += nri;
1092
1093     /* Update outer/inner flops */
1094
1095     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*387);
1096 }
1097 /*
1098  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_F_avx_128_fma_double
1099  * Electrostatics interaction: ReactionField
1100  * VdW interaction:            CubicSplineTable
1101  * Geometry:                   Water3-Water3
1102  * Calculate force/pot:        Force
1103  */
1104 void
1105 nb_kernel_ElecRFCut_VdwCSTab_GeomW3W3_F_avx_128_fma_double
1106                     (t_nblist                    * gmx_restrict       nlist,
1107                      rvec                        * gmx_restrict          xx,
1108                      rvec                        * gmx_restrict          ff,
1109                      t_forcerec                  * gmx_restrict          fr,
1110                      t_mdatoms                   * gmx_restrict     mdatoms,
1111                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1112                      t_nrnb                      * gmx_restrict        nrnb)
1113 {
1114     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1115      * just 0 for non-waters.
1116      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1117      * jnr indices corresponding to data put in the four positions in the SIMD register.
1118      */
1119     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1120     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1121     int              jnrA,jnrB;
1122     int              j_coord_offsetA,j_coord_offsetB;
1123     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1124     real             rcutoff_scalar;
1125     real             *shiftvec,*fshift,*x,*f;
1126     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1127     int              vdwioffset0;
1128     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1129     int              vdwioffset1;
1130     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1131     int              vdwioffset2;
1132     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1133     int              vdwjidx0A,vdwjidx0B;
1134     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1135     int              vdwjidx1A,vdwjidx1B;
1136     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1137     int              vdwjidx2A,vdwjidx2B;
1138     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1139     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1140     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1141     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1142     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1143     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1144     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1145     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1146     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1147     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1148     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1149     real             *charge;
1150     int              nvdwtype;
1151     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1152     int              *vdwtype;
1153     real             *vdwparam;
1154     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1155     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1156     __m128i          vfitab;
1157     __m128i          ifour       = _mm_set1_epi32(4);
1158     __m128d          rt,vfeps,vftabscale,Y,F,G,H,Heps,Fp,VV,FF,twovfeps;
1159     real             *vftab;
1160     __m128d          dummy_mask,cutoff_mask;
1161     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1162     __m128d          one     = _mm_set1_pd(1.0);
1163     __m128d          two     = _mm_set1_pd(2.0);
1164     x                = xx[0];
1165     f                = ff[0];
1166
1167     nri              = nlist->nri;
1168     iinr             = nlist->iinr;
1169     jindex           = nlist->jindex;
1170     jjnr             = nlist->jjnr;
1171     shiftidx         = nlist->shift;
1172     gid              = nlist->gid;
1173     shiftvec         = fr->shift_vec[0];
1174     fshift           = fr->fshift[0];
1175     facel            = _mm_set1_pd(fr->epsfac);
1176     charge           = mdatoms->chargeA;
1177     krf              = _mm_set1_pd(fr->ic->k_rf);
1178     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1179     crf              = _mm_set1_pd(fr->ic->c_rf);
1180     nvdwtype         = fr->ntype;
1181     vdwparam         = fr->nbfp;
1182     vdwtype          = mdatoms->typeA;
1183
1184     vftab            = kernel_data->table_vdw->data;
1185     vftabscale       = _mm_set1_pd(kernel_data->table_vdw->scale);
1186
1187     /* Setup water-specific parameters */
1188     inr              = nlist->iinr[0];
1189     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
1190     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1191     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1192     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1193
1194     jq0              = _mm_set1_pd(charge[inr+0]);
1195     jq1              = _mm_set1_pd(charge[inr+1]);
1196     jq2              = _mm_set1_pd(charge[inr+2]);
1197     vdwjidx0A        = 2*vdwtype[inr+0];
1198     qq00             = _mm_mul_pd(iq0,jq0);
1199     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1200     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1201     qq01             = _mm_mul_pd(iq0,jq1);
1202     qq02             = _mm_mul_pd(iq0,jq2);
1203     qq10             = _mm_mul_pd(iq1,jq0);
1204     qq11             = _mm_mul_pd(iq1,jq1);
1205     qq12             = _mm_mul_pd(iq1,jq2);
1206     qq20             = _mm_mul_pd(iq2,jq0);
1207     qq21             = _mm_mul_pd(iq2,jq1);
1208     qq22             = _mm_mul_pd(iq2,jq2);
1209
1210     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1211     rcutoff_scalar   = fr->rcoulomb;
1212     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1213     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1214
1215     /* Avoid stupid compiler warnings */
1216     jnrA = jnrB = 0;
1217     j_coord_offsetA = 0;
1218     j_coord_offsetB = 0;
1219
1220     outeriter        = 0;
1221     inneriter        = 0;
1222
1223     /* Start outer loop over neighborlists */
1224     for(iidx=0; iidx<nri; iidx++)
1225     {
1226         /* Load shift vector for this list */
1227         i_shift_offset   = DIM*shiftidx[iidx];
1228
1229         /* Load limits for loop over neighbors */
1230         j_index_start    = jindex[iidx];
1231         j_index_end      = jindex[iidx+1];
1232
1233         /* Get outer coordinate index */
1234         inr              = iinr[iidx];
1235         i_coord_offset   = DIM*inr;
1236
1237         /* Load i particle coords and add shift vector */
1238         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1239                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1240
1241         fix0             = _mm_setzero_pd();
1242         fiy0             = _mm_setzero_pd();
1243         fiz0             = _mm_setzero_pd();
1244         fix1             = _mm_setzero_pd();
1245         fiy1             = _mm_setzero_pd();
1246         fiz1             = _mm_setzero_pd();
1247         fix2             = _mm_setzero_pd();
1248         fiy2             = _mm_setzero_pd();
1249         fiz2             = _mm_setzero_pd();
1250
1251         /* Start inner kernel loop */
1252         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1253         {
1254
1255             /* Get j neighbor index, and coordinate index */
1256             jnrA             = jjnr[jidx];
1257             jnrB             = jjnr[jidx+1];
1258             j_coord_offsetA  = DIM*jnrA;
1259             j_coord_offsetB  = DIM*jnrB;
1260
1261             /* load j atom coordinates */
1262             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1263                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1264
1265             /* Calculate displacement vector */
1266             dx00             = _mm_sub_pd(ix0,jx0);
1267             dy00             = _mm_sub_pd(iy0,jy0);
1268             dz00             = _mm_sub_pd(iz0,jz0);
1269             dx01             = _mm_sub_pd(ix0,jx1);
1270             dy01             = _mm_sub_pd(iy0,jy1);
1271             dz01             = _mm_sub_pd(iz0,jz1);
1272             dx02             = _mm_sub_pd(ix0,jx2);
1273             dy02             = _mm_sub_pd(iy0,jy2);
1274             dz02             = _mm_sub_pd(iz0,jz2);
1275             dx10             = _mm_sub_pd(ix1,jx0);
1276             dy10             = _mm_sub_pd(iy1,jy0);
1277             dz10             = _mm_sub_pd(iz1,jz0);
1278             dx11             = _mm_sub_pd(ix1,jx1);
1279             dy11             = _mm_sub_pd(iy1,jy1);
1280             dz11             = _mm_sub_pd(iz1,jz1);
1281             dx12             = _mm_sub_pd(ix1,jx2);
1282             dy12             = _mm_sub_pd(iy1,jy2);
1283             dz12             = _mm_sub_pd(iz1,jz2);
1284             dx20             = _mm_sub_pd(ix2,jx0);
1285             dy20             = _mm_sub_pd(iy2,jy0);
1286             dz20             = _mm_sub_pd(iz2,jz0);
1287             dx21             = _mm_sub_pd(ix2,jx1);
1288             dy21             = _mm_sub_pd(iy2,jy1);
1289             dz21             = _mm_sub_pd(iz2,jz1);
1290             dx22             = _mm_sub_pd(ix2,jx2);
1291             dy22             = _mm_sub_pd(iy2,jy2);
1292             dz22             = _mm_sub_pd(iz2,jz2);
1293
1294             /* Calculate squared distance and things based on it */
1295             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1296             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1297             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1298             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1299             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1300             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1301             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1302             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1303             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1304
1305             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1306             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1307             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1308             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1309             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1310             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1311             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1312             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1313             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1314
1315             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1316             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1317             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1318             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1319             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1320             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1321             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1322             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1323             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1324
1325             fjx0             = _mm_setzero_pd();
1326             fjy0             = _mm_setzero_pd();
1327             fjz0             = _mm_setzero_pd();
1328             fjx1             = _mm_setzero_pd();
1329             fjy1             = _mm_setzero_pd();
1330             fjz1             = _mm_setzero_pd();
1331             fjx2             = _mm_setzero_pd();
1332             fjy2             = _mm_setzero_pd();
1333             fjz2             = _mm_setzero_pd();
1334
1335             /**************************
1336              * CALCULATE INTERACTIONS *
1337              **************************/
1338
1339             if (gmx_mm_any_lt(rsq00,rcutoff2))
1340             {
1341
1342             r00              = _mm_mul_pd(rsq00,rinv00);
1343
1344             /* Calculate table index by multiplying r with table scale and truncate to integer */
1345             rt               = _mm_mul_pd(r00,vftabscale);
1346             vfitab           = _mm_cvttpd_epi32(rt);
1347 #ifdef __XOP__
1348             vfeps            = _mm_frcz_pd(rt);
1349 #else
1350             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
1351 #endif
1352             twovfeps         = _mm_add_pd(vfeps,vfeps);
1353             vfitab           = _mm_slli_epi32(vfitab,3);
1354
1355             /* REACTION-FIELD ELECTROSTATICS */
1356             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
1357
1358             /* CUBIC SPLINE TABLE DISPERSION */
1359             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1360             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
1361             GMX_MM_TRANSPOSE2_PD(Y,F);
1362             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1363             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
1364             GMX_MM_TRANSPOSE2_PD(G,H);
1365             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1366             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1367             fvdw6            = _mm_mul_pd(c6_00,FF);
1368
1369             /* CUBIC SPLINE TABLE REPULSION */
1370             vfitab           = _mm_add_epi32(vfitab,ifour);
1371             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1372             F                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) );
1373             GMX_MM_TRANSPOSE2_PD(Y,F);
1374             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1375             H                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,1) +2);
1376             GMX_MM_TRANSPOSE2_PD(G,H);
1377             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1378             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1379             fvdw12           = _mm_mul_pd(c12_00,FF);
1380             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1381
1382             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1383
1384             fscal            = _mm_add_pd(felec,fvdw);
1385
1386             fscal            = _mm_and_pd(fscal,cutoff_mask);
1387
1388             /* Update vectorial force */
1389             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1390             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1391             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1392
1393             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1394             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1395             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1396
1397             }
1398
1399             /**************************
1400              * CALCULATE INTERACTIONS *
1401              **************************/
1402
1403             if (gmx_mm_any_lt(rsq01,rcutoff2))
1404             {
1405
1406             /* REACTION-FIELD ELECTROSTATICS */
1407             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
1408
1409             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
1410
1411             fscal            = felec;
1412
1413             fscal            = _mm_and_pd(fscal,cutoff_mask);
1414
1415             /* Update vectorial force */
1416             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1417             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1418             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1419
1420             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1421             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1422             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1423
1424             }
1425
1426             /**************************
1427              * CALCULATE INTERACTIONS *
1428              **************************/
1429
1430             if (gmx_mm_any_lt(rsq02,rcutoff2))
1431             {
1432
1433             /* REACTION-FIELD ELECTROSTATICS */
1434             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
1435
1436             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
1437
1438             fscal            = felec;
1439
1440             fscal            = _mm_and_pd(fscal,cutoff_mask);
1441
1442             /* Update vectorial force */
1443             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1444             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1445             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1446
1447             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1448             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1449             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1450
1451             }
1452
1453             /**************************
1454              * CALCULATE INTERACTIONS *
1455              **************************/
1456
1457             if (gmx_mm_any_lt(rsq10,rcutoff2))
1458             {
1459
1460             /* REACTION-FIELD ELECTROSTATICS */
1461             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
1462
1463             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
1464
1465             fscal            = felec;
1466
1467             fscal            = _mm_and_pd(fscal,cutoff_mask);
1468
1469             /* Update vectorial force */
1470             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1471             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1472             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1473
1474             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1475             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1476             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1477
1478             }
1479
1480             /**************************
1481              * CALCULATE INTERACTIONS *
1482              **************************/
1483
1484             if (gmx_mm_any_lt(rsq11,rcutoff2))
1485             {
1486
1487             /* REACTION-FIELD ELECTROSTATICS */
1488             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1489
1490             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1491
1492             fscal            = felec;
1493
1494             fscal            = _mm_and_pd(fscal,cutoff_mask);
1495
1496             /* Update vectorial force */
1497             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1498             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1499             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1500
1501             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1502             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1503             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1504
1505             }
1506
1507             /**************************
1508              * CALCULATE INTERACTIONS *
1509              **************************/
1510
1511             if (gmx_mm_any_lt(rsq12,rcutoff2))
1512             {
1513
1514             /* REACTION-FIELD ELECTROSTATICS */
1515             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1516
1517             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1518
1519             fscal            = felec;
1520
1521             fscal            = _mm_and_pd(fscal,cutoff_mask);
1522
1523             /* Update vectorial force */
1524             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1525             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1526             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1527
1528             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1529             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1530             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1531
1532             }
1533
1534             /**************************
1535              * CALCULATE INTERACTIONS *
1536              **************************/
1537
1538             if (gmx_mm_any_lt(rsq20,rcutoff2))
1539             {
1540
1541             /* REACTION-FIELD ELECTROSTATICS */
1542             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
1543
1544             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
1545
1546             fscal            = felec;
1547
1548             fscal            = _mm_and_pd(fscal,cutoff_mask);
1549
1550             /* Update vectorial force */
1551             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1552             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1553             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1554
1555             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1556             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1557             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1558
1559             }
1560
1561             /**************************
1562              * CALCULATE INTERACTIONS *
1563              **************************/
1564
1565             if (gmx_mm_any_lt(rsq21,rcutoff2))
1566             {
1567
1568             /* REACTION-FIELD ELECTROSTATICS */
1569             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1570
1571             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1572
1573             fscal            = felec;
1574
1575             fscal            = _mm_and_pd(fscal,cutoff_mask);
1576
1577             /* Update vectorial force */
1578             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1579             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1580             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1581
1582             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1583             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1584             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1585
1586             }
1587
1588             /**************************
1589              * CALCULATE INTERACTIONS *
1590              **************************/
1591
1592             if (gmx_mm_any_lt(rsq22,rcutoff2))
1593             {
1594
1595             /* REACTION-FIELD ELECTROSTATICS */
1596             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1597
1598             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1599
1600             fscal            = felec;
1601
1602             fscal            = _mm_and_pd(fscal,cutoff_mask);
1603
1604             /* Update vectorial force */
1605             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1606             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1607             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1608
1609             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1610             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1611             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1612
1613             }
1614
1615             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1616
1617             /* Inner loop uses 324 flops */
1618         }
1619
1620         if(jidx<j_index_end)
1621         {
1622
1623             jnrA             = jjnr[jidx];
1624             j_coord_offsetA  = DIM*jnrA;
1625
1626             /* load j atom coordinates */
1627             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1628                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1629
1630             /* Calculate displacement vector */
1631             dx00             = _mm_sub_pd(ix0,jx0);
1632             dy00             = _mm_sub_pd(iy0,jy0);
1633             dz00             = _mm_sub_pd(iz0,jz0);
1634             dx01             = _mm_sub_pd(ix0,jx1);
1635             dy01             = _mm_sub_pd(iy0,jy1);
1636             dz01             = _mm_sub_pd(iz0,jz1);
1637             dx02             = _mm_sub_pd(ix0,jx2);
1638             dy02             = _mm_sub_pd(iy0,jy2);
1639             dz02             = _mm_sub_pd(iz0,jz2);
1640             dx10             = _mm_sub_pd(ix1,jx0);
1641             dy10             = _mm_sub_pd(iy1,jy0);
1642             dz10             = _mm_sub_pd(iz1,jz0);
1643             dx11             = _mm_sub_pd(ix1,jx1);
1644             dy11             = _mm_sub_pd(iy1,jy1);
1645             dz11             = _mm_sub_pd(iz1,jz1);
1646             dx12             = _mm_sub_pd(ix1,jx2);
1647             dy12             = _mm_sub_pd(iy1,jy2);
1648             dz12             = _mm_sub_pd(iz1,jz2);
1649             dx20             = _mm_sub_pd(ix2,jx0);
1650             dy20             = _mm_sub_pd(iy2,jy0);
1651             dz20             = _mm_sub_pd(iz2,jz0);
1652             dx21             = _mm_sub_pd(ix2,jx1);
1653             dy21             = _mm_sub_pd(iy2,jy1);
1654             dz21             = _mm_sub_pd(iz2,jz1);
1655             dx22             = _mm_sub_pd(ix2,jx2);
1656             dy22             = _mm_sub_pd(iy2,jy2);
1657             dz22             = _mm_sub_pd(iz2,jz2);
1658
1659             /* Calculate squared distance and things based on it */
1660             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1661             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1662             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1663             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1664             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1665             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1666             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1667             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1668             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1669
1670             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1671             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1672             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1673             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1674             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1675             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1676             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1677             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1678             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1679
1680             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1681             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1682             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1683             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1684             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1685             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1686             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1687             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1688             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1689
1690             fjx0             = _mm_setzero_pd();
1691             fjy0             = _mm_setzero_pd();
1692             fjz0             = _mm_setzero_pd();
1693             fjx1             = _mm_setzero_pd();
1694             fjy1             = _mm_setzero_pd();
1695             fjz1             = _mm_setzero_pd();
1696             fjx2             = _mm_setzero_pd();
1697             fjy2             = _mm_setzero_pd();
1698             fjz2             = _mm_setzero_pd();
1699
1700             /**************************
1701              * CALCULATE INTERACTIONS *
1702              **************************/
1703
1704             if (gmx_mm_any_lt(rsq00,rcutoff2))
1705             {
1706
1707             r00              = _mm_mul_pd(rsq00,rinv00);
1708
1709             /* Calculate table index by multiplying r with table scale and truncate to integer */
1710             rt               = _mm_mul_pd(r00,vftabscale);
1711             vfitab           = _mm_cvttpd_epi32(rt);
1712 #ifdef __XOP__
1713             vfeps            = _mm_frcz_pd(rt);
1714 #else
1715             vfeps            = _mm_sub_pd(rt,_mm_round_pd(rt, _MM_FROUND_FLOOR));
1716 #endif
1717             twovfeps         = _mm_add_pd(vfeps,vfeps);
1718             vfitab           = _mm_slli_epi32(vfitab,3);
1719
1720             /* REACTION-FIELD ELECTROSTATICS */
1721             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
1722
1723             /* CUBIC SPLINE TABLE DISPERSION */
1724             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1725             F                = _mm_setzero_pd();
1726             GMX_MM_TRANSPOSE2_PD(Y,F);
1727             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1728             H                = _mm_setzero_pd();
1729             GMX_MM_TRANSPOSE2_PD(G,H);
1730             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1731             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1732             fvdw6            = _mm_mul_pd(c6_00,FF);
1733
1734             /* CUBIC SPLINE TABLE REPULSION */
1735             vfitab           = _mm_add_epi32(vfitab,ifour);
1736             Y                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) );
1737             F                = _mm_setzero_pd();
1738             GMX_MM_TRANSPOSE2_PD(Y,F);
1739             G                = _mm_load_pd( vftab + _mm_extract_epi32(vfitab,0) +2);
1740             H                = _mm_setzero_pd();
1741             GMX_MM_TRANSPOSE2_PD(G,H);
1742             Fp               = _mm_macc_pd(vfeps,_mm_macc_pd(H,vfeps,G),F);
1743             FF               = _mm_macc_pd(vfeps,_mm_macc_pd(twovfeps,H,G),Fp);
1744             fvdw12           = _mm_mul_pd(c12_00,FF);
1745             fvdw             = _mm_xor_pd(signbit,_mm_mul_pd(_mm_add_pd(fvdw6,fvdw12),_mm_mul_pd(vftabscale,rinv00)));
1746
1747             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1748
1749             fscal            = _mm_add_pd(felec,fvdw);
1750
1751             fscal            = _mm_and_pd(fscal,cutoff_mask);
1752
1753             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1754
1755             /* Update vectorial force */
1756             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1757             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1758             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1759
1760             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1761             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1762             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1763
1764             }
1765
1766             /**************************
1767              * CALCULATE INTERACTIONS *
1768              **************************/
1769
1770             if (gmx_mm_any_lt(rsq01,rcutoff2))
1771             {
1772
1773             /* REACTION-FIELD ELECTROSTATICS */
1774             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
1775
1776             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
1777
1778             fscal            = felec;
1779
1780             fscal            = _mm_and_pd(fscal,cutoff_mask);
1781
1782             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1783
1784             /* Update vectorial force */
1785             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1786             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1787             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1788
1789             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1790             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1791             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1792
1793             }
1794
1795             /**************************
1796              * CALCULATE INTERACTIONS *
1797              **************************/
1798
1799             if (gmx_mm_any_lt(rsq02,rcutoff2))
1800             {
1801
1802             /* REACTION-FIELD ELECTROSTATICS */
1803             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
1804
1805             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
1806
1807             fscal            = felec;
1808
1809             fscal            = _mm_and_pd(fscal,cutoff_mask);
1810
1811             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1812
1813             /* Update vectorial force */
1814             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1815             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1816             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1817
1818             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1819             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1820             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1821
1822             }
1823
1824             /**************************
1825              * CALCULATE INTERACTIONS *
1826              **************************/
1827
1828             if (gmx_mm_any_lt(rsq10,rcutoff2))
1829             {
1830
1831             /* REACTION-FIELD ELECTROSTATICS */
1832             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
1833
1834             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
1835
1836             fscal            = felec;
1837
1838             fscal            = _mm_and_pd(fscal,cutoff_mask);
1839
1840             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1841
1842             /* Update vectorial force */
1843             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1844             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1845             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1846
1847             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1848             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1849             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1850
1851             }
1852
1853             /**************************
1854              * CALCULATE INTERACTIONS *
1855              **************************/
1856
1857             if (gmx_mm_any_lt(rsq11,rcutoff2))
1858             {
1859
1860             /* REACTION-FIELD ELECTROSTATICS */
1861             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1862
1863             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1864
1865             fscal            = felec;
1866
1867             fscal            = _mm_and_pd(fscal,cutoff_mask);
1868
1869             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1870
1871             /* Update vectorial force */
1872             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1873             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1874             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1875
1876             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1877             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1878             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1879
1880             }
1881
1882             /**************************
1883              * CALCULATE INTERACTIONS *
1884              **************************/
1885
1886             if (gmx_mm_any_lt(rsq12,rcutoff2))
1887             {
1888
1889             /* REACTION-FIELD ELECTROSTATICS */
1890             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1891
1892             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1893
1894             fscal            = felec;
1895
1896             fscal            = _mm_and_pd(fscal,cutoff_mask);
1897
1898             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1899
1900             /* Update vectorial force */
1901             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1902             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1903             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1904
1905             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1906             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1907             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1908
1909             }
1910
1911             /**************************
1912              * CALCULATE INTERACTIONS *
1913              **************************/
1914
1915             if (gmx_mm_any_lt(rsq20,rcutoff2))
1916             {
1917
1918             /* REACTION-FIELD ELECTROSTATICS */
1919             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
1920
1921             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
1922
1923             fscal            = felec;
1924
1925             fscal            = _mm_and_pd(fscal,cutoff_mask);
1926
1927             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1928
1929             /* Update vectorial force */
1930             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1931             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1932             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1933
1934             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1935             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1936             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1937
1938             }
1939
1940             /**************************
1941              * CALCULATE INTERACTIONS *
1942              **************************/
1943
1944             if (gmx_mm_any_lt(rsq21,rcutoff2))
1945             {
1946
1947             /* REACTION-FIELD ELECTROSTATICS */
1948             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1949
1950             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1951
1952             fscal            = felec;
1953
1954             fscal            = _mm_and_pd(fscal,cutoff_mask);
1955
1956             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1957
1958             /* Update vectorial force */
1959             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1960             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1961             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1962
1963             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1964             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1965             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1966
1967             }
1968
1969             /**************************
1970              * CALCULATE INTERACTIONS *
1971              **************************/
1972
1973             if (gmx_mm_any_lt(rsq22,rcutoff2))
1974             {
1975
1976             /* REACTION-FIELD ELECTROSTATICS */
1977             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1978
1979             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1980
1981             fscal            = felec;
1982
1983             fscal            = _mm_and_pd(fscal,cutoff_mask);
1984
1985             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1986
1987             /* Update vectorial force */
1988             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1989             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1990             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1991
1992             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1993             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1994             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1995
1996             }
1997
1998             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1999
2000             /* Inner loop uses 324 flops */
2001         }
2002
2003         /* End of innermost loop */
2004
2005         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
2006                                               f+i_coord_offset,fshift+i_shift_offset);
2007
2008         /* Increment number of inner iterations */
2009         inneriter                  += j_index_end - j_index_start;
2010
2011         /* Outer loop uses 18 flops */
2012     }
2013
2014     /* Increment number of outer iterations */
2015     outeriter        += nri;
2016
2017     /* Update outer/inner flops */
2018
2019     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*324);
2020 }