src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecRFCut_VdwLJSw_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 "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_double.h"
  48 #include "kernelutil_x86_avx_128_fma_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomW3W3_VF_avx_128_fma_double
  52  * Electrostatics interaction: ReactionField
  53  * VdW interaction:            LennardJones
  54  * Geometry:                   Water3-Water3
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecRFCut_VdwLJSw_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     __m128d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 110     real             rswitch_scalar,d_scalar;
 111     __m128d          dummy_mask,cutoff_mask;
 112     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 113     __m128d          one     = _mm_set1_pd(1.0);
 114     __m128d          two     = _mm_set1_pd(2.0);
 115     x                = xx[0];
 116     f                = ff[0];
 117
 118     nri              = nlist->nri;
 119     iinr             = nlist->iinr;
 120     jindex           = nlist->jindex;
 121     jjnr             = nlist->jjnr;
 122     shiftidx         = nlist->shift;
 123     gid              = nlist->gid;
 124     shiftvec         = fr->shift_vec[0];
 125     fshift           = fr->fshift[0];
 126     facel            = _mm_set1_pd(fr->epsfac);
 127     charge           = mdatoms->chargeA;
 128     krf              = _mm_set1_pd(fr->ic->k_rf);
 129     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 130     crf              = _mm_set1_pd(fr->ic->c_rf);
 131     nvdwtype         = fr->ntype;
 132     vdwparam         = fr->nbfp;
 133     vdwtype          = mdatoms->typeA;
 134
 135     /* Setup water-specific parameters */
 136     inr              = nlist->iinr[0];
 137     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 138     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 139     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 140     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 141
 142     jq0              = _mm_set1_pd(charge[inr+0]);
 143     jq1              = _mm_set1_pd(charge[inr+1]);
 144     jq2              = _mm_set1_pd(charge[inr+2]);
 145     vdwjidx0A        = 2*vdwtype[inr+0];
 146     qq00             = _mm_mul_pd(iq0,jq0);
 147     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 148     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 149     qq01             = _mm_mul_pd(iq0,jq1);
 150     qq02             = _mm_mul_pd(iq0,jq2);
 151     qq10             = _mm_mul_pd(iq1,jq0);
 152     qq11             = _mm_mul_pd(iq1,jq1);
 153     qq12             = _mm_mul_pd(iq1,jq2);
 154     qq20             = _mm_mul_pd(iq2,jq0);
 155     qq21             = _mm_mul_pd(iq2,jq1);
 156     qq22             = _mm_mul_pd(iq2,jq2);
 157
 158     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 159     rcutoff_scalar   = fr->rcoulomb;
 160     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 161     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 162
 163     rswitch_scalar   = fr->rvdw_switch;
 164     rswitch          = _mm_set1_pd(rswitch_scalar);
 165     /* Setup switch parameters */
 166     d_scalar         = rcutoff_scalar-rswitch_scalar;
 167     d                = _mm_set1_pd(d_scalar);
 168     swV3             = _mm_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
 169     swV4             = _mm_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 170     swV5             = _mm_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 171     swF2             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
 172     swF3             = _mm_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 173     swF4             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 174
 175     /* Avoid stupid compiler warnings */
 176     jnrA = jnrB = 0;
 177     j_coord_offsetA = 0;
 178     j_coord_offsetB = 0;
 179
 180     outeriter        = 0;
 181     inneriter        = 0;
 182
 183     /* Start outer loop over neighborlists */
 184     for(iidx=0; iidx<nri; iidx++)
 185     {
 186         /* Load shift vector for this list */
 187         i_shift_offset   = DIM*shiftidx[iidx];
 188
 189         /* Load limits for loop over neighbors */
 190         j_index_start    = jindex[iidx];
 191         j_index_end      = jindex[iidx+1];
 192
 193         /* Get outer coordinate index */
 194         inr              = iinr[iidx];
 195         i_coord_offset   = DIM*inr;
 196
 197         /* Load i particle coords and add shift vector */
 198         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 199                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 200
 201         fix0             = _mm_setzero_pd();
 202         fiy0             = _mm_setzero_pd();
 203         fiz0             = _mm_setzero_pd();
 204         fix1             = _mm_setzero_pd();
 205         fiy1             = _mm_setzero_pd();
 206         fiz1             = _mm_setzero_pd();
 207         fix2             = _mm_setzero_pd();
 208         fiy2             = _mm_setzero_pd();
 209         fiz2             = _mm_setzero_pd();
 210
 211         /* Reset potential sums */
 212         velecsum         = _mm_setzero_pd();
 213         vvdwsum          = _mm_setzero_pd();
 214
 215         /* Start inner kernel loop */
 216         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 217         {
 218
 219             /* Get j neighbor index, and coordinate index */
 220             jnrA             = jjnr[jidx];
 221             jnrB             = jjnr[jidx+1];
 222             j_coord_offsetA  = DIM*jnrA;
 223             j_coord_offsetB  = DIM*jnrB;
 224
 225             /* load j atom coordinates */
 226             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 227                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 228
 229             /* Calculate displacement vector */
 230             dx00             = _mm_sub_pd(ix0,jx0);
 231             dy00             = _mm_sub_pd(iy0,jy0);
 232             dz00             = _mm_sub_pd(iz0,jz0);
 233             dx01             = _mm_sub_pd(ix0,jx1);
 234             dy01             = _mm_sub_pd(iy0,jy1);
 235             dz01             = _mm_sub_pd(iz0,jz1);
 236             dx02             = _mm_sub_pd(ix0,jx2);
 237             dy02             = _mm_sub_pd(iy0,jy2);
 238             dz02             = _mm_sub_pd(iz0,jz2);
 239             dx10             = _mm_sub_pd(ix1,jx0);
 240             dy10             = _mm_sub_pd(iy1,jy0);
 241             dz10             = _mm_sub_pd(iz1,jz0);
 242             dx11             = _mm_sub_pd(ix1,jx1);
 243             dy11             = _mm_sub_pd(iy1,jy1);
 244             dz11             = _mm_sub_pd(iz1,jz1);
 245             dx12             = _mm_sub_pd(ix1,jx2);
 246             dy12             = _mm_sub_pd(iy1,jy2);
 247             dz12             = _mm_sub_pd(iz1,jz2);
 248             dx20             = _mm_sub_pd(ix2,jx0);
 249             dy20             = _mm_sub_pd(iy2,jy0);
 250             dz20             = _mm_sub_pd(iz2,jz0);
 251             dx21             = _mm_sub_pd(ix2,jx1);
 252             dy21             = _mm_sub_pd(iy2,jy1);
 253             dz21             = _mm_sub_pd(iz2,jz1);
 254             dx22             = _mm_sub_pd(ix2,jx2);
 255             dy22             = _mm_sub_pd(iy2,jy2);
 256             dz22             = _mm_sub_pd(iz2,jz2);
 257
 258             /* Calculate squared distance and things based on it */
 259             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 260             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 261             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 262             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 263             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 264             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 265             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 266             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 267             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 268
 269             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 270             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 271             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 272             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 273             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 274             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 275             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 276             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 277             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 278
 279             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 280             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 281             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 282             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 283             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 284             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 285             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 286             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 287             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 288
 289             fjx0             = _mm_setzero_pd();
 290             fjy0             = _mm_setzero_pd();
 291             fjz0             = _mm_setzero_pd();
 292             fjx1             = _mm_setzero_pd();
 293             fjy1             = _mm_setzero_pd();
 294             fjz1             = _mm_setzero_pd();
 295             fjx2             = _mm_setzero_pd();
 296             fjy2             = _mm_setzero_pd();
 297             fjz2             = _mm_setzero_pd();
 298
 299             /**************************
 300              * CALCULATE INTERACTIONS *
 301              **************************/
 302
 303             if (gmx_mm_any_lt(rsq00,rcutoff2))
 304             {
 305
 306             r00              = _mm_mul_pd(rsq00,rinv00);
 307
 308             /* REACTION-FIELD ELECTROSTATICS */
 309             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
 310             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 311
 312             /* LENNARD-JONES DISPERSION/REPULSION */
 313
 314             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 315             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 316             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 317             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 318             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 319
 320             d                = _mm_sub_pd(r00,rswitch);
 321             d                = _mm_max_pd(d,_mm_setzero_pd());
 322             d2               = _mm_mul_pd(d,d);
 323             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 324
 325             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 326
 327             /* Evaluate switch function */
 328             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 329             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 330             vvdw             = _mm_mul_pd(vvdw,sw);
 331             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 332
 333             /* Update potential sum for this i atom from the interaction with this j atom. */
 334             velec            = _mm_and_pd(velec,cutoff_mask);
 335             velecsum         = _mm_add_pd(velecsum,velec);
 336             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 337             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 338
 339             fscal            = _mm_add_pd(felec,fvdw);
 340
 341             fscal            = _mm_and_pd(fscal,cutoff_mask);
 342
 343             /* Update vectorial force */
 344             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 345             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 346             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 347
 348             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 349             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 350             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 351
 352             }
 353
 354             /**************************
 355              * CALCULATE INTERACTIONS *
 356              **************************/
 357
 358             if (gmx_mm_any_lt(rsq01,rcutoff2))
 359             {
 360
 361             /* REACTION-FIELD ELECTROSTATICS */
 362             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_macc_pd(krf,rsq01,rinv01),crf));
 363             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
 364
 365             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
 366
 367             /* Update potential sum for this i atom from the interaction with this j atom. */
 368             velec            = _mm_and_pd(velec,cutoff_mask);
 369             velecsum         = _mm_add_pd(velecsum,velec);
 370
 371             fscal            = felec;
 372
 373             fscal            = _mm_and_pd(fscal,cutoff_mask);
 374
 375             /* Update vectorial force */
 376             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 377             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 378             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 379
 380             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 381             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 382             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 383
 384             }
 385
 386             /**************************
 387              * CALCULATE INTERACTIONS *
 388              **************************/
 389
 390             if (gmx_mm_any_lt(rsq02,rcutoff2))
 391             {
 392
 393             /* REACTION-FIELD ELECTROSTATICS */
 394             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_macc_pd(krf,rsq02,rinv02),crf));
 395             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
 396
 397             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
 398
 399             /* Update potential sum for this i atom from the interaction with this j atom. */
 400             velec            = _mm_and_pd(velec,cutoff_mask);
 401             velecsum         = _mm_add_pd(velecsum,velec);
 402
 403             fscal            = felec;
 404
 405             fscal            = _mm_and_pd(fscal,cutoff_mask);
 406
 407             /* Update vectorial force */
 408             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 409             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 410             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 411
 412             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 413             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 414             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 415
 416             }
 417
 418             /**************************
 419              * CALCULATE INTERACTIONS *
 420              **************************/
 421
 422             if (gmx_mm_any_lt(rsq10,rcutoff2))
 423             {
 424
 425             /* REACTION-FIELD ELECTROSTATICS */
 426             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
 427             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 428
 429             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
 430
 431             /* Update potential sum for this i atom from the interaction with this j atom. */
 432             velec            = _mm_and_pd(velec,cutoff_mask);
 433             velecsum         = _mm_add_pd(velecsum,velec);
 434
 435             fscal            = felec;
 436
 437             fscal            = _mm_and_pd(fscal,cutoff_mask);
 438
 439             /* Update vectorial force */
 440             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 441             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 442             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 443
 444             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 445             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 446             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 447
 448             }
 449
 450             /**************************
 451              * CALCULATE INTERACTIONS *
 452              **************************/
 453
 454             if (gmx_mm_any_lt(rsq11,rcutoff2))
 455             {
 456
 457             /* REACTION-FIELD ELECTROSTATICS */
 458             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 459             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 460
 461             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 462
 463             /* Update potential sum for this i atom from the interaction with this j atom. */
 464             velec            = _mm_and_pd(velec,cutoff_mask);
 465             velecsum         = _mm_add_pd(velecsum,velec);
 466
 467             fscal            = felec;
 468
 469             fscal            = _mm_and_pd(fscal,cutoff_mask);
 470
 471             /* Update vectorial force */
 472             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 473             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 474             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 475
 476             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 477             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 478             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 479
 480             }
 481
 482             /**************************
 483              * CALCULATE INTERACTIONS *
 484              **************************/
 485
 486             if (gmx_mm_any_lt(rsq12,rcutoff2))
 487             {
 488
 489             /* REACTION-FIELD ELECTROSTATICS */
 490             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 491             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 492
 493             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 494
 495             /* Update potential sum for this i atom from the interaction with this j atom. */
 496             velec            = _mm_and_pd(velec,cutoff_mask);
 497             velecsum         = _mm_add_pd(velecsum,velec);
 498
 499             fscal            = felec;
 500
 501             fscal            = _mm_and_pd(fscal,cutoff_mask);
 502
 503             /* Update vectorial force */
 504             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 505             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 506             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 507
 508             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 509             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 510             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 511
 512             }
 513
 514             /**************************
 515              * CALCULATE INTERACTIONS *
 516              **************************/
 517
 518             if (gmx_mm_any_lt(rsq20,rcutoff2))
 519             {
 520
 521             /* REACTION-FIELD ELECTROSTATICS */
 522             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
 523             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 524
 525             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
 526
 527             /* Update potential sum for this i atom from the interaction with this j atom. */
 528             velec            = _mm_and_pd(velec,cutoff_mask);
 529             velecsum         = _mm_add_pd(velecsum,velec);
 530
 531             fscal            = felec;
 532
 533             fscal            = _mm_and_pd(fscal,cutoff_mask);
 534
 535             /* Update vectorial force */
 536             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 537             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 538             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 539
 540             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 541             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 542             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 543
 544             }
 545
 546             /**************************
 547              * CALCULATE INTERACTIONS *
 548              **************************/
 549
 550             if (gmx_mm_any_lt(rsq21,rcutoff2))
 551             {
 552
 553             /* REACTION-FIELD ELECTROSTATICS */
 554             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 555             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 556
 557             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 558
 559             /* Update potential sum for this i atom from the interaction with this j atom. */
 560             velec            = _mm_and_pd(velec,cutoff_mask);
 561             velecsum         = _mm_add_pd(velecsum,velec);
 562
 563             fscal            = felec;
 564
 565             fscal            = _mm_and_pd(fscal,cutoff_mask);
 566
 567             /* Update vectorial force */
 568             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 569             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 570             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 571
 572             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 573             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 574             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 575
 576             }
 577
 578             /**************************
 579              * CALCULATE INTERACTIONS *
 580              **************************/
 581
 582             if (gmx_mm_any_lt(rsq22,rcutoff2))
 583             {
 584
 585             /* REACTION-FIELD ELECTROSTATICS */
 586             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 587             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 588
 589             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 590
 591             /* Update potential sum for this i atom from the interaction with this j atom. */
 592             velec            = _mm_and_pd(velec,cutoff_mask);
 593             velecsum         = _mm_add_pd(velecsum,velec);
 594
 595             fscal            = felec;
 596
 597             fscal            = _mm_and_pd(fscal,cutoff_mask);
 598
 599             /* Update vectorial force */
 600             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 601             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 602             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 603
 604             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 605             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 606             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 607
 608             }
 609
 610             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 611
 612             /* Inner loop uses 385 flops */
 613         }
 614
 615         if(jidx<j_index_end)
 616         {
 617
 618             jnrA             = jjnr[jidx];
 619             j_coord_offsetA  = DIM*jnrA;
 620
 621             /* load j atom coordinates */
 622             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 623                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 624
 625             /* Calculate displacement vector */
 626             dx00             = _mm_sub_pd(ix0,jx0);
 627             dy00             = _mm_sub_pd(iy0,jy0);
 628             dz00             = _mm_sub_pd(iz0,jz0);
 629             dx01             = _mm_sub_pd(ix0,jx1);
 630             dy01             = _mm_sub_pd(iy0,jy1);
 631             dz01             = _mm_sub_pd(iz0,jz1);
 632             dx02             = _mm_sub_pd(ix0,jx2);
 633             dy02             = _mm_sub_pd(iy0,jy2);
 634             dz02             = _mm_sub_pd(iz0,jz2);
 635             dx10             = _mm_sub_pd(ix1,jx0);
 636             dy10             = _mm_sub_pd(iy1,jy0);
 637             dz10             = _mm_sub_pd(iz1,jz0);
 638             dx11             = _mm_sub_pd(ix1,jx1);
 639             dy11             = _mm_sub_pd(iy1,jy1);
 640             dz11             = _mm_sub_pd(iz1,jz1);
 641             dx12             = _mm_sub_pd(ix1,jx2);
 642             dy12             = _mm_sub_pd(iy1,jy2);
 643             dz12             = _mm_sub_pd(iz1,jz2);
 644             dx20             = _mm_sub_pd(ix2,jx0);
 645             dy20             = _mm_sub_pd(iy2,jy0);
 646             dz20             = _mm_sub_pd(iz2,jz0);
 647             dx21             = _mm_sub_pd(ix2,jx1);
 648             dy21             = _mm_sub_pd(iy2,jy1);
 649             dz21             = _mm_sub_pd(iz2,jz1);
 650             dx22             = _mm_sub_pd(ix2,jx2);
 651             dy22             = _mm_sub_pd(iy2,jy2);
 652             dz22             = _mm_sub_pd(iz2,jz2);
 653
 654             /* Calculate squared distance and things based on it */
 655             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 656             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 657             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 658             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 659             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 660             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 661             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 662             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 663             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 664
 665             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 666             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 667             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 668             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 669             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 670             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 671             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 672             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 673             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 674
 675             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 676             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 677             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 678             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 679             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 680             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 681             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 682             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 683             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 684
 685             fjx0             = _mm_setzero_pd();
 686             fjy0             = _mm_setzero_pd();
 687             fjz0             = _mm_setzero_pd();
 688             fjx1             = _mm_setzero_pd();
 689             fjy1             = _mm_setzero_pd();
 690             fjz1             = _mm_setzero_pd();
 691             fjx2             = _mm_setzero_pd();
 692             fjy2             = _mm_setzero_pd();
 693             fjz2             = _mm_setzero_pd();
 694
 695             /**************************
 696              * CALCULATE INTERACTIONS *
 697              **************************/
 698
 699             if (gmx_mm_any_lt(rsq00,rcutoff2))
 700             {
 701
 702             r00              = _mm_mul_pd(rsq00,rinv00);
 703
 704             /* REACTION-FIELD ELECTROSTATICS */
 705             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
 706             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 707
 708             /* LENNARD-JONES DISPERSION/REPULSION */
 709
 710             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 711             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 712             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 713             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 714             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 715
 716             d                = _mm_sub_pd(r00,rswitch);
 717             d                = _mm_max_pd(d,_mm_setzero_pd());
 718             d2               = _mm_mul_pd(d,d);
 719             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
 720
 721             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
 722
 723             /* Evaluate switch function */
 724             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 725             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
 726             vvdw             = _mm_mul_pd(vvdw,sw);
 727             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 728
 729             /* Update potential sum for this i atom from the interaction with this j atom. */
 730             velec            = _mm_and_pd(velec,cutoff_mask);
 731             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 732             velecsum         = _mm_add_pd(velecsum,velec);
 733             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 734             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 735             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 736
 737             fscal            = _mm_add_pd(felec,fvdw);
 738
 739             fscal            = _mm_and_pd(fscal,cutoff_mask);
 740
 741             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 742
 743             /* Update vectorial force */
 744             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 745             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 746             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 747
 748             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 749             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 750             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 751
 752             }
 753
 754             /**************************
 755              * CALCULATE INTERACTIONS *
 756              **************************/
 757
 758             if (gmx_mm_any_lt(rsq01,rcutoff2))
 759             {
 760
 761             /* REACTION-FIELD ELECTROSTATICS */
 762             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_macc_pd(krf,rsq01,rinv01),crf));
 763             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
 764
 765             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
 766
 767             /* Update potential sum for this i atom from the interaction with this j atom. */
 768             velec            = _mm_and_pd(velec,cutoff_mask);
 769             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 770             velecsum         = _mm_add_pd(velecsum,velec);
 771
 772             fscal            = felec;
 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(dx01,fscal,fix0);
 780             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 781             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 782
 783             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 784             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 785             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 786
 787             }
 788
 789             /**************************
 790              * CALCULATE INTERACTIONS *
 791              **************************/
 792
 793             if (gmx_mm_any_lt(rsq02,rcutoff2))
 794             {
 795
 796             /* REACTION-FIELD ELECTROSTATICS */
 797             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_macc_pd(krf,rsq02,rinv02),crf));
 798             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
 799
 800             cutoff_mask      = _mm_cmplt_pd(rsq02,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(dx02,fscal,fix0);
 815             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 816             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 817
 818             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 819             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 820             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 821
 822             }
 823
 824             /**************************
 825              * CALCULATE INTERACTIONS *
 826              **************************/
 827
 828             if (gmx_mm_any_lt(rsq10,rcutoff2))
 829             {
 830
 831             /* REACTION-FIELD ELECTROSTATICS */
 832             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
 833             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 834
 835             cutoff_mask      = _mm_cmplt_pd(rsq10,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             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 850             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 851             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 852
 853             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 854             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 855             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 856
 857             }
 858
 859             /**************************
 860              * CALCULATE INTERACTIONS *
 861              **************************/
 862
 863             if (gmx_mm_any_lt(rsq11,rcutoff2))
 864             {
 865
 866             /* REACTION-FIELD ELECTROSTATICS */
 867             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 868             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 869
 870             cutoff_mask      = _mm_cmplt_pd(rsq11,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(dx11,fscal,fix1);
 885             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 886             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 887
 888             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 889             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 890             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 891
 892             }
 893
 894             /**************************
 895              * CALCULATE INTERACTIONS *
 896              **************************/
 897
 898             if (gmx_mm_any_lt(rsq12,rcutoff2))
 899             {
 900
 901             /* REACTION-FIELD ELECTROSTATICS */
 902             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 903             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 904
 905             cutoff_mask      = _mm_cmplt_pd(rsq12,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(dx12,fscal,fix1);
 920             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 921             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 922
 923             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 924             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 925             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 926
 927             }
 928
 929             /**************************
 930              * CALCULATE INTERACTIONS *
 931              **************************/
 932
 933             if (gmx_mm_any_lt(rsq20,rcutoff2))
 934             {
 935
 936             /* REACTION-FIELD ELECTROSTATICS */
 937             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
 938             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 939
 940             cutoff_mask      = _mm_cmplt_pd(rsq20,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             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 955             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 956             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 957
 958             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 959             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 960             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 961
 962             }
 963
 964             /**************************
 965              * CALCULATE INTERACTIONS *
 966              **************************/
 967
 968             if (gmx_mm_any_lt(rsq21,rcutoff2))
 969             {
 970
 971             /* REACTION-FIELD ELECTROSTATICS */
 972             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 973             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 974
 975             cutoff_mask      = _mm_cmplt_pd(rsq21,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(dx21,fscal,fix2);
 990             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 991             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 992
 993             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 994             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 995             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 996
 997             }
 998
 999             /**************************
1000              * CALCULATE INTERACTIONS *
1001              **************************/
1002
1003             if (gmx_mm_any_lt(rsq22,rcutoff2))
1004             {
1005
1006             /* REACTION-FIELD ELECTROSTATICS */
1007             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
1008             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1009
1010             cutoff_mask      = _mm_cmplt_pd(rsq22,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(dx22,fscal,fix2);
1025             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1026             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1027
1028             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1029             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1030             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1031
1032             }
1033
1034             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1035
1036             /* Inner loop uses 385 flops */
1037         }
1038
1039         /* End of innermost loop */
1040
1041         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1042                                               f+i_coord_offset,fshift+i_shift_offset);
1043
1044         ggid                        = gid[iidx];
1045         /* Update potential energies */
1046         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1047         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
1048
1049         /* Increment number of inner iterations */
1050         inneriter                  += j_index_end - j_index_start;
1051
1052         /* Outer loop uses 20 flops */
1053     }
1054
1055     /* Increment number of outer iterations */
1056     outeriter        += nri;
1057
1058     /* Update outer/inner flops */
1059
1060     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*385);
1061 }
1062 /*
1063  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomW3W3_F_avx_128_fma_double
1064  * Electrostatics interaction: ReactionField
1065  * VdW interaction:            LennardJones
1066  * Geometry:                   Water3-Water3
1067  * Calculate force/pot:        Force
1068  */
1069 void
1070 nb_kernel_ElecRFCut_VdwLJSw_GeomW3W3_F_avx_128_fma_double
1071                     (t_nblist                    * gmx_restrict       nlist,
1072                      rvec                        * gmx_restrict          xx,
1073                      rvec                        * gmx_restrict          ff,
1074                      t_forcerec                  * gmx_restrict          fr,
1075                      t_mdatoms                   * gmx_restrict     mdatoms,
1076                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1077                      t_nrnb                      * gmx_restrict        nrnb)
1078 {
1079     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1080      * just 0 for non-waters.
1081      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1082      * jnr indices corresponding to data put in the four positions in the SIMD register.
1083      */
1084     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1085     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1086     int              jnrA,jnrB;
1087     int              j_coord_offsetA,j_coord_offsetB;
1088     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1089     real             rcutoff_scalar;
1090     real             *shiftvec,*fshift,*x,*f;
1091     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1092     int              vdwioffset0;
1093     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1094     int              vdwioffset1;
1095     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1096     int              vdwioffset2;
1097     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1098     int              vdwjidx0A,vdwjidx0B;
1099     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1100     int              vdwjidx1A,vdwjidx1B;
1101     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1102     int              vdwjidx2A,vdwjidx2B;
1103     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1104     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1105     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1106     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1107     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1108     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1109     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1110     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1111     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1112     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1113     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1114     real             *charge;
1115     int              nvdwtype;
1116     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1117     int              *vdwtype;
1118     real             *vdwparam;
1119     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1120     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1121     __m128d          rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
1122     real             rswitch_scalar,d_scalar;
1123     __m128d          dummy_mask,cutoff_mask;
1124     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1125     __m128d          one     = _mm_set1_pd(1.0);
1126     __m128d          two     = _mm_set1_pd(2.0);
1127     x                = xx[0];
1128     f                = ff[0];
1129
1130     nri              = nlist->nri;
1131     iinr             = nlist->iinr;
1132     jindex           = nlist->jindex;
1133     jjnr             = nlist->jjnr;
1134     shiftidx         = nlist->shift;
1135     gid              = nlist->gid;
1136     shiftvec         = fr->shift_vec[0];
1137     fshift           = fr->fshift[0];
1138     facel            = _mm_set1_pd(fr->epsfac);
1139     charge           = mdatoms->chargeA;
1140     krf              = _mm_set1_pd(fr->ic->k_rf);
1141     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1142     crf              = _mm_set1_pd(fr->ic->c_rf);
1143     nvdwtype         = fr->ntype;
1144     vdwparam         = fr->nbfp;
1145     vdwtype          = mdatoms->typeA;
1146
1147     /* Setup water-specific parameters */
1148     inr              = nlist->iinr[0];
1149     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
1150     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1151     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1152     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1153
1154     jq0              = _mm_set1_pd(charge[inr+0]);
1155     jq1              = _mm_set1_pd(charge[inr+1]);
1156     jq2              = _mm_set1_pd(charge[inr+2]);
1157     vdwjidx0A        = 2*vdwtype[inr+0];
1158     qq00             = _mm_mul_pd(iq0,jq0);
1159     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1160     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1161     qq01             = _mm_mul_pd(iq0,jq1);
1162     qq02             = _mm_mul_pd(iq0,jq2);
1163     qq10             = _mm_mul_pd(iq1,jq0);
1164     qq11             = _mm_mul_pd(iq1,jq1);
1165     qq12             = _mm_mul_pd(iq1,jq2);
1166     qq20             = _mm_mul_pd(iq2,jq0);
1167     qq21             = _mm_mul_pd(iq2,jq1);
1168     qq22             = _mm_mul_pd(iq2,jq2);
1169
1170     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1171     rcutoff_scalar   = fr->rcoulomb;
1172     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1173     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1174
1175     rswitch_scalar   = fr->rvdw_switch;
1176     rswitch          = _mm_set1_pd(rswitch_scalar);
1177     /* Setup switch parameters */
1178     d_scalar         = rcutoff_scalar-rswitch_scalar;
1179     d                = _mm_set1_pd(d_scalar);
1180     swV3             = _mm_set1_pd(-10.0/(d_scalar*d_scalar*d_scalar));
1181     swV4             = _mm_set1_pd( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1182     swV5             = _mm_set1_pd( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1183     swF2             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar));
1184     swF3             = _mm_set1_pd( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1185     swF4             = _mm_set1_pd(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1186
1187     /* Avoid stupid compiler warnings */
1188     jnrA = jnrB = 0;
1189     j_coord_offsetA = 0;
1190     j_coord_offsetB = 0;
1191
1192     outeriter        = 0;
1193     inneriter        = 0;
1194
1195     /* Start outer loop over neighborlists */
1196     for(iidx=0; iidx<nri; iidx++)
1197     {
1198         /* Load shift vector for this list */
1199         i_shift_offset   = DIM*shiftidx[iidx];
1200
1201         /* Load limits for loop over neighbors */
1202         j_index_start    = jindex[iidx];
1203         j_index_end      = jindex[iidx+1];
1204
1205         /* Get outer coordinate index */
1206         inr              = iinr[iidx];
1207         i_coord_offset   = DIM*inr;
1208
1209         /* Load i particle coords and add shift vector */
1210         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1211                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1212
1213         fix0             = _mm_setzero_pd();
1214         fiy0             = _mm_setzero_pd();
1215         fiz0             = _mm_setzero_pd();
1216         fix1             = _mm_setzero_pd();
1217         fiy1             = _mm_setzero_pd();
1218         fiz1             = _mm_setzero_pd();
1219         fix2             = _mm_setzero_pd();
1220         fiy2             = _mm_setzero_pd();
1221         fiz2             = _mm_setzero_pd();
1222
1223         /* Start inner kernel loop */
1224         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1225         {
1226
1227             /* Get j neighbor index, and coordinate index */
1228             jnrA             = jjnr[jidx];
1229             jnrB             = jjnr[jidx+1];
1230             j_coord_offsetA  = DIM*jnrA;
1231             j_coord_offsetB  = DIM*jnrB;
1232
1233             /* load j atom coordinates */
1234             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1235                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1236
1237             /* Calculate displacement vector */
1238             dx00             = _mm_sub_pd(ix0,jx0);
1239             dy00             = _mm_sub_pd(iy0,jy0);
1240             dz00             = _mm_sub_pd(iz0,jz0);
1241             dx01             = _mm_sub_pd(ix0,jx1);
1242             dy01             = _mm_sub_pd(iy0,jy1);
1243             dz01             = _mm_sub_pd(iz0,jz1);
1244             dx02             = _mm_sub_pd(ix0,jx2);
1245             dy02             = _mm_sub_pd(iy0,jy2);
1246             dz02             = _mm_sub_pd(iz0,jz2);
1247             dx10             = _mm_sub_pd(ix1,jx0);
1248             dy10             = _mm_sub_pd(iy1,jy0);
1249             dz10             = _mm_sub_pd(iz1,jz0);
1250             dx11             = _mm_sub_pd(ix1,jx1);
1251             dy11             = _mm_sub_pd(iy1,jy1);
1252             dz11             = _mm_sub_pd(iz1,jz1);
1253             dx12             = _mm_sub_pd(ix1,jx2);
1254             dy12             = _mm_sub_pd(iy1,jy2);
1255             dz12             = _mm_sub_pd(iz1,jz2);
1256             dx20             = _mm_sub_pd(ix2,jx0);
1257             dy20             = _mm_sub_pd(iy2,jy0);
1258             dz20             = _mm_sub_pd(iz2,jz0);
1259             dx21             = _mm_sub_pd(ix2,jx1);
1260             dy21             = _mm_sub_pd(iy2,jy1);
1261             dz21             = _mm_sub_pd(iz2,jz1);
1262             dx22             = _mm_sub_pd(ix2,jx2);
1263             dy22             = _mm_sub_pd(iy2,jy2);
1264             dz22             = _mm_sub_pd(iz2,jz2);
1265
1266             /* Calculate squared distance and things based on it */
1267             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1268             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1269             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1270             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1271             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1272             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1273             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1274             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1275             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1276
1277             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1278             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1279             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1280             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1281             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1282             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1283             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1284             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1285             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1286
1287             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1288             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1289             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1290             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1291             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1292             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1293             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1294             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1295             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1296
1297             fjx0             = _mm_setzero_pd();
1298             fjy0             = _mm_setzero_pd();
1299             fjz0             = _mm_setzero_pd();
1300             fjx1             = _mm_setzero_pd();
1301             fjy1             = _mm_setzero_pd();
1302             fjz1             = _mm_setzero_pd();
1303             fjx2             = _mm_setzero_pd();
1304             fjy2             = _mm_setzero_pd();
1305             fjz2             = _mm_setzero_pd();
1306
1307             /**************************
1308              * CALCULATE INTERACTIONS *
1309              **************************/
1310
1311             if (gmx_mm_any_lt(rsq00,rcutoff2))
1312             {
1313
1314             r00              = _mm_mul_pd(rsq00,rinv00);
1315
1316             /* REACTION-FIELD ELECTROSTATICS */
1317             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
1318
1319             /* LENNARD-JONES DISPERSION/REPULSION */
1320
1321             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1322             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
1323             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
1324             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
1325             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
1326
1327             d                = _mm_sub_pd(r00,rswitch);
1328             d                = _mm_max_pd(d,_mm_setzero_pd());
1329             d2               = _mm_mul_pd(d,d);
1330             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
1331
1332             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
1333
1334             /* Evaluate switch function */
1335             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1336             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
1337             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1338
1339             fscal            = _mm_add_pd(felec,fvdw);
1340
1341             fscal            = _mm_and_pd(fscal,cutoff_mask);
1342
1343             /* Update vectorial force */
1344             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1345             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1346             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1347
1348             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1349             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1350             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1351
1352             }
1353
1354             /**************************
1355              * CALCULATE INTERACTIONS *
1356              **************************/
1357
1358             if (gmx_mm_any_lt(rsq01,rcutoff2))
1359             {
1360
1361             /* REACTION-FIELD ELECTROSTATICS */
1362             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
1363
1364             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
1365
1366             fscal            = felec;
1367
1368             fscal            = _mm_and_pd(fscal,cutoff_mask);
1369
1370             /* Update vectorial force */
1371             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1372             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1373             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1374
1375             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1376             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1377             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1378
1379             }
1380
1381             /**************************
1382              * CALCULATE INTERACTIONS *
1383              **************************/
1384
1385             if (gmx_mm_any_lt(rsq02,rcutoff2))
1386             {
1387
1388             /* REACTION-FIELD ELECTROSTATICS */
1389             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
1390
1391             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
1392
1393             fscal            = felec;
1394
1395             fscal            = _mm_and_pd(fscal,cutoff_mask);
1396
1397             /* Update vectorial force */
1398             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1399             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1400             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1401
1402             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1403             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1404             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1405
1406             }
1407
1408             /**************************
1409              * CALCULATE INTERACTIONS *
1410              **************************/
1411
1412             if (gmx_mm_any_lt(rsq10,rcutoff2))
1413             {
1414
1415             /* REACTION-FIELD ELECTROSTATICS */
1416             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
1417
1418             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
1419
1420             fscal            = felec;
1421
1422             fscal            = _mm_and_pd(fscal,cutoff_mask);
1423
1424             /* Update vectorial force */
1425             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1426             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1427             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1428
1429             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1430             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1431             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1432
1433             }
1434
1435             /**************************
1436              * CALCULATE INTERACTIONS *
1437              **************************/
1438
1439             if (gmx_mm_any_lt(rsq11,rcutoff2))
1440             {
1441
1442             /* REACTION-FIELD ELECTROSTATICS */
1443             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1444
1445             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1446
1447             fscal            = felec;
1448
1449             fscal            = _mm_and_pd(fscal,cutoff_mask);
1450
1451             /* Update vectorial force */
1452             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1453             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1454             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1455
1456             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1457             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1458             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1459
1460             }
1461
1462             /**************************
1463              * CALCULATE INTERACTIONS *
1464              **************************/
1465
1466             if (gmx_mm_any_lt(rsq12,rcutoff2))
1467             {
1468
1469             /* REACTION-FIELD ELECTROSTATICS */
1470             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1471
1472             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1473
1474             fscal            = felec;
1475
1476             fscal            = _mm_and_pd(fscal,cutoff_mask);
1477
1478             /* Update vectorial force */
1479             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1480             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1481             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1482
1483             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1484             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1485             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1486
1487             }
1488
1489             /**************************
1490              * CALCULATE INTERACTIONS *
1491              **************************/
1492
1493             if (gmx_mm_any_lt(rsq20,rcutoff2))
1494             {
1495
1496             /* REACTION-FIELD ELECTROSTATICS */
1497             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
1498
1499             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
1500
1501             fscal            = felec;
1502
1503             fscal            = _mm_and_pd(fscal,cutoff_mask);
1504
1505             /* Update vectorial force */
1506             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1507             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1508             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1509
1510             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1511             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1512             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1513
1514             }
1515
1516             /**************************
1517              * CALCULATE INTERACTIONS *
1518              **************************/
1519
1520             if (gmx_mm_any_lt(rsq21,rcutoff2))
1521             {
1522
1523             /* REACTION-FIELD ELECTROSTATICS */
1524             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1525
1526             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1527
1528             fscal            = felec;
1529
1530             fscal            = _mm_and_pd(fscal,cutoff_mask);
1531
1532             /* Update vectorial force */
1533             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1534             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1535             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1536
1537             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1538             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1539             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1540
1541             }
1542
1543             /**************************
1544              * CALCULATE INTERACTIONS *
1545              **************************/
1546
1547             if (gmx_mm_any_lt(rsq22,rcutoff2))
1548             {
1549
1550             /* REACTION-FIELD ELECTROSTATICS */
1551             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1552
1553             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1554
1555             fscal            = felec;
1556
1557             fscal            = _mm_and_pd(fscal,cutoff_mask);
1558
1559             /* Update vectorial force */
1560             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1561             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1562             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1563
1564             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1565             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1566             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1567
1568             }
1569
1570             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1571
1572             /* Inner loop uses 328 flops */
1573         }
1574
1575         if(jidx<j_index_end)
1576         {
1577
1578             jnrA             = jjnr[jidx];
1579             j_coord_offsetA  = DIM*jnrA;
1580
1581             /* load j atom coordinates */
1582             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1583                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1584
1585             /* Calculate displacement vector */
1586             dx00             = _mm_sub_pd(ix0,jx0);
1587             dy00             = _mm_sub_pd(iy0,jy0);
1588             dz00             = _mm_sub_pd(iz0,jz0);
1589             dx01             = _mm_sub_pd(ix0,jx1);
1590             dy01             = _mm_sub_pd(iy0,jy1);
1591             dz01             = _mm_sub_pd(iz0,jz1);
1592             dx02             = _mm_sub_pd(ix0,jx2);
1593             dy02             = _mm_sub_pd(iy0,jy2);
1594             dz02             = _mm_sub_pd(iz0,jz2);
1595             dx10             = _mm_sub_pd(ix1,jx0);
1596             dy10             = _mm_sub_pd(iy1,jy0);
1597             dz10             = _mm_sub_pd(iz1,jz0);
1598             dx11             = _mm_sub_pd(ix1,jx1);
1599             dy11             = _mm_sub_pd(iy1,jy1);
1600             dz11             = _mm_sub_pd(iz1,jz1);
1601             dx12             = _mm_sub_pd(ix1,jx2);
1602             dy12             = _mm_sub_pd(iy1,jy2);
1603             dz12             = _mm_sub_pd(iz1,jz2);
1604             dx20             = _mm_sub_pd(ix2,jx0);
1605             dy20             = _mm_sub_pd(iy2,jy0);
1606             dz20             = _mm_sub_pd(iz2,jz0);
1607             dx21             = _mm_sub_pd(ix2,jx1);
1608             dy21             = _mm_sub_pd(iy2,jy1);
1609             dz21             = _mm_sub_pd(iz2,jz1);
1610             dx22             = _mm_sub_pd(ix2,jx2);
1611             dy22             = _mm_sub_pd(iy2,jy2);
1612             dz22             = _mm_sub_pd(iz2,jz2);
1613
1614             /* Calculate squared distance and things based on it */
1615             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1616             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1617             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1618             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1619             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1620             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1621             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1622             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1623             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1624
1625             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1626             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1627             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1628             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1629             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1630             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1631             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1632             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1633             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1634
1635             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1636             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1637             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1638             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1639             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1640             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1641             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1642             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1643             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1644
1645             fjx0             = _mm_setzero_pd();
1646             fjy0             = _mm_setzero_pd();
1647             fjz0             = _mm_setzero_pd();
1648             fjx1             = _mm_setzero_pd();
1649             fjy1             = _mm_setzero_pd();
1650             fjz1             = _mm_setzero_pd();
1651             fjx2             = _mm_setzero_pd();
1652             fjy2             = _mm_setzero_pd();
1653             fjz2             = _mm_setzero_pd();
1654
1655             /**************************
1656              * CALCULATE INTERACTIONS *
1657              **************************/
1658
1659             if (gmx_mm_any_lt(rsq00,rcutoff2))
1660             {
1661
1662             r00              = _mm_mul_pd(rsq00,rinv00);
1663
1664             /* REACTION-FIELD ELECTROSTATICS */
1665             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
1666
1667             /* LENNARD-JONES DISPERSION/REPULSION */
1668
1669             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1670             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
1671             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
1672             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
1673             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
1674
1675             d                = _mm_sub_pd(r00,rswitch);
1676             d                = _mm_max_pd(d,_mm_setzero_pd());
1677             d2               = _mm_mul_pd(d,d);
1678             sw               = _mm_add_pd(one,_mm_mul_pd(d2,_mm_mul_pd(d,_mm_macc_pd(d,_mm_macc_pd(d,swV5,swV4),swV3))));
1679
1680             dsw              = _mm_mul_pd(d2,_mm_macc_pd(d,_mm_macc_pd(d,swF4,swF3),swF2));
1681
1682             /* Evaluate switch function */
1683             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1684             fvdw             = _mm_msub_pd( fvdw,sw , _mm_mul_pd(rinv00,_mm_mul_pd(vvdw,dsw)) );
1685             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1686
1687             fscal            = _mm_add_pd(felec,fvdw);
1688
1689             fscal            = _mm_and_pd(fscal,cutoff_mask);
1690
1691             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1692
1693             /* Update vectorial force */
1694             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1695             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1696             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1697
1698             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1699             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1700             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1701
1702             }
1703
1704             /**************************
1705              * CALCULATE INTERACTIONS *
1706              **************************/
1707
1708             if (gmx_mm_any_lt(rsq01,rcutoff2))
1709             {
1710
1711             /* REACTION-FIELD ELECTROSTATICS */
1712             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
1713
1714             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
1715
1716             fscal            = felec;
1717
1718             fscal            = _mm_and_pd(fscal,cutoff_mask);
1719
1720             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1721
1722             /* Update vectorial force */
1723             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1724             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1725             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1726
1727             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1728             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1729             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1730
1731             }
1732
1733             /**************************
1734              * CALCULATE INTERACTIONS *
1735              **************************/
1736
1737             if (gmx_mm_any_lt(rsq02,rcutoff2))
1738             {
1739
1740             /* REACTION-FIELD ELECTROSTATICS */
1741             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
1742
1743             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
1744
1745             fscal            = felec;
1746
1747             fscal            = _mm_and_pd(fscal,cutoff_mask);
1748
1749             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1750
1751             /* Update vectorial force */
1752             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1753             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1754             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1755
1756             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1757             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1758             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1759
1760             }
1761
1762             /**************************
1763              * CALCULATE INTERACTIONS *
1764              **************************/
1765
1766             if (gmx_mm_any_lt(rsq10,rcutoff2))
1767             {
1768
1769             /* REACTION-FIELD ELECTROSTATICS */
1770             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
1771
1772             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
1773
1774             fscal            = felec;
1775
1776             fscal            = _mm_and_pd(fscal,cutoff_mask);
1777
1778             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1779
1780             /* Update vectorial force */
1781             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1782             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1783             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1784
1785             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1786             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1787             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1788
1789             }
1790
1791             /**************************
1792              * CALCULATE INTERACTIONS *
1793              **************************/
1794
1795             if (gmx_mm_any_lt(rsq11,rcutoff2))
1796             {
1797
1798             /* REACTION-FIELD ELECTROSTATICS */
1799             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1800
1801             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1802
1803             fscal            = felec;
1804
1805             fscal            = _mm_and_pd(fscal,cutoff_mask);
1806
1807             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1808
1809             /* Update vectorial force */
1810             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1811             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1812             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1813
1814             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1815             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1816             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1817
1818             }
1819
1820             /**************************
1821              * CALCULATE INTERACTIONS *
1822              **************************/
1823
1824             if (gmx_mm_any_lt(rsq12,rcutoff2))
1825             {
1826
1827             /* REACTION-FIELD ELECTROSTATICS */
1828             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1829
1830             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1831
1832             fscal            = felec;
1833
1834             fscal            = _mm_and_pd(fscal,cutoff_mask);
1835
1836             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1837
1838             /* Update vectorial force */
1839             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1840             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1841             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1842
1843             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1844             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1845             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1846
1847             }
1848
1849             /**************************
1850              * CALCULATE INTERACTIONS *
1851              **************************/
1852
1853             if (gmx_mm_any_lt(rsq20,rcutoff2))
1854             {
1855
1856             /* REACTION-FIELD ELECTROSTATICS */
1857             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
1858
1859             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
1860
1861             fscal            = felec;
1862
1863             fscal            = _mm_and_pd(fscal,cutoff_mask);
1864
1865             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1866
1867             /* Update vectorial force */
1868             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1869             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1870             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1871
1872             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1873             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1874             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1875
1876             }
1877
1878             /**************************
1879              * CALCULATE INTERACTIONS *
1880              **************************/
1881
1882             if (gmx_mm_any_lt(rsq21,rcutoff2))
1883             {
1884
1885             /* REACTION-FIELD ELECTROSTATICS */
1886             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1887
1888             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1889
1890             fscal            = felec;
1891
1892             fscal            = _mm_and_pd(fscal,cutoff_mask);
1893
1894             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1895
1896             /* Update vectorial force */
1897             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1898             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1899             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1900
1901             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1902             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1903             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1904
1905             }
1906
1907             /**************************
1908              * CALCULATE INTERACTIONS *
1909              **************************/
1910
1911             if (gmx_mm_any_lt(rsq22,rcutoff2))
1912             {
1913
1914             /* REACTION-FIELD ELECTROSTATICS */
1915             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1916
1917             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1918
1919             fscal            = felec;
1920
1921             fscal            = _mm_and_pd(fscal,cutoff_mask);
1922
1923             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1924
1925             /* Update vectorial force */
1926             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1927             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1928             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1929
1930             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1931             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1932             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1933
1934             }
1935
1936             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1937
1938             /* Inner loop uses 328 flops */
1939         }
1940
1941         /* End of innermost loop */
1942
1943         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1944                                               f+i_coord_offset,fshift+i_shift_offset);
1945
1946         /* Increment number of inner iterations */
1947         inneriter                  += j_index_end - j_index_start;
1948
1949         /* Outer loop uses 18 flops */
1950     }
1951
1952     /* Increment number of outer iterations */
1953     outeriter        += nri;
1954
1955     /* Update outer/inner flops */
1956
1957     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*328);
1958 }