src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecRFCut_VdwLJSw_GeomW3W3_avx_128_fma_single.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_single kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_avx_128_fma_single.h"
  50 #include "kernelutil_x86_avx_128_fma_single.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomW3W3_VF_avx_128_fma_single
  54  * Electrostatics interaction: ReactionField
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Water3-Water3
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecRFCut_VdwLJSw_GeomW3W3_VF_avx_128_fma_single
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB,jnrC,jnrD;
  77     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  78     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  79     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  80     real             rcutoff_scalar;
  81     real             *shiftvec,*fshift,*x,*f;
  82     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  83     real             scratch[4*DIM];
  84     __m128           fscal,rcutoff,rcutoff2,jidxall;
  85     int              vdwioffset0;
  86     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  87     int              vdwioffset1;
  88     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  89     int              vdwioffset2;
  90     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  91     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  92     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  93     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  94     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  95     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  96     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  97     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  98     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  99     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
 100     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
 101     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 102     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 103     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
 104     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 105     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 106     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 107     real             *charge;
 108     int              nvdwtype;
 109     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 110     int              *vdwtype;
 111     real             *vdwparam;
 112     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 113     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 114     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
 115     real             rswitch_scalar,d_scalar;
 116     __m128           dummy_mask,cutoff_mask;
 117     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 118     __m128           one     = _mm_set1_ps(1.0);
 119     __m128           two     = _mm_set1_ps(2.0);
 120     x                = xx[0];
 121     f                = ff[0];
 122
 123     nri              = nlist->nri;
 124     iinr             = nlist->iinr;
 125     jindex           = nlist->jindex;
 126     jjnr             = nlist->jjnr;
 127     shiftidx         = nlist->shift;
 128     gid              = nlist->gid;
 129     shiftvec         = fr->shift_vec[0];
 130     fshift           = fr->fshift[0];
 131     facel            = _mm_set1_ps(fr->epsfac);
 132     charge           = mdatoms->chargeA;
 133     krf              = _mm_set1_ps(fr->ic->k_rf);
 134     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 135     crf              = _mm_set1_ps(fr->ic->c_rf);
 136     nvdwtype         = fr->ntype;
 137     vdwparam         = fr->nbfp;
 138     vdwtype          = mdatoms->typeA;
 139
 140     /* Setup water-specific parameters */
 141     inr              = nlist->iinr[0];
 142     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
 143     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 144     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 145     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 146
 147     jq0              = _mm_set1_ps(charge[inr+0]);
 148     jq1              = _mm_set1_ps(charge[inr+1]);
 149     jq2              = _mm_set1_ps(charge[inr+2]);
 150     vdwjidx0A        = 2*vdwtype[inr+0];
 151     qq00             = _mm_mul_ps(iq0,jq0);
 152     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 153     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 154     qq01             = _mm_mul_ps(iq0,jq1);
 155     qq02             = _mm_mul_ps(iq0,jq2);
 156     qq10             = _mm_mul_ps(iq1,jq0);
 157     qq11             = _mm_mul_ps(iq1,jq1);
 158     qq12             = _mm_mul_ps(iq1,jq2);
 159     qq20             = _mm_mul_ps(iq2,jq0);
 160     qq21             = _mm_mul_ps(iq2,jq1);
 161     qq22             = _mm_mul_ps(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_ps(rcutoff_scalar);
 166     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 167
 168     rswitch_scalar   = fr->rvdw_switch;
 169     rswitch          = _mm_set1_ps(rswitch_scalar);
 170     /* Setup switch parameters */
 171     d_scalar         = rcutoff_scalar-rswitch_scalar;
 172     d                = _mm_set1_ps(d_scalar);
 173     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
 174     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 175     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 176     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
 177     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
 178     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
 179
 180     /* Avoid stupid compiler warnings */
 181     jnrA = jnrB = jnrC = jnrD = 0;
 182     j_coord_offsetA = 0;
 183     j_coord_offsetB = 0;
 184     j_coord_offsetC = 0;
 185     j_coord_offsetD = 0;
 186
 187     outeriter        = 0;
 188     inneriter        = 0;
 189
 190     for(iidx=0;iidx<4*DIM;iidx++)
 191     {
 192         scratch[iidx] = 0.0;
 193     }
 194
 195     /* Start outer loop over neighborlists */
 196     for(iidx=0; iidx<nri; iidx++)
 197     {
 198         /* Load shift vector for this list */
 199         i_shift_offset   = DIM*shiftidx[iidx];
 200
 201         /* Load limits for loop over neighbors */
 202         j_index_start    = jindex[iidx];
 203         j_index_end      = jindex[iidx+1];
 204
 205         /* Get outer coordinate index */
 206         inr              = iinr[iidx];
 207         i_coord_offset   = DIM*inr;
 208
 209         /* Load i particle coords and add shift vector */
 210         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 211                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 212
 213         fix0             = _mm_setzero_ps();
 214         fiy0             = _mm_setzero_ps();
 215         fiz0             = _mm_setzero_ps();
 216         fix1             = _mm_setzero_ps();
 217         fiy1             = _mm_setzero_ps();
 218         fiz1             = _mm_setzero_ps();
 219         fix2             = _mm_setzero_ps();
 220         fiy2             = _mm_setzero_ps();
 221         fiz2             = _mm_setzero_ps();
 222
 223         /* Reset potential sums */
 224         velecsum         = _mm_setzero_ps();
 225         vvdwsum          = _mm_setzero_ps();
 226
 227         /* Start inner kernel loop */
 228         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 229         {
 230
 231             /* Get j neighbor index, and coordinate index */
 232             jnrA             = jjnr[jidx];
 233             jnrB             = jjnr[jidx+1];
 234             jnrC             = jjnr[jidx+2];
 235             jnrD             = jjnr[jidx+3];
 236             j_coord_offsetA  = DIM*jnrA;
 237             j_coord_offsetB  = DIM*jnrB;
 238             j_coord_offsetC  = DIM*jnrC;
 239             j_coord_offsetD  = DIM*jnrD;
 240
 241             /* load j atom coordinates */
 242             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 243                                               x+j_coord_offsetC,x+j_coord_offsetD,
 244                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 245
 246             /* Calculate displacement vector */
 247             dx00             = _mm_sub_ps(ix0,jx0);
 248             dy00             = _mm_sub_ps(iy0,jy0);
 249             dz00             = _mm_sub_ps(iz0,jz0);
 250             dx01             = _mm_sub_ps(ix0,jx1);
 251             dy01             = _mm_sub_ps(iy0,jy1);
 252             dz01             = _mm_sub_ps(iz0,jz1);
 253             dx02             = _mm_sub_ps(ix0,jx2);
 254             dy02             = _mm_sub_ps(iy0,jy2);
 255             dz02             = _mm_sub_ps(iz0,jz2);
 256             dx10             = _mm_sub_ps(ix1,jx0);
 257             dy10             = _mm_sub_ps(iy1,jy0);
 258             dz10             = _mm_sub_ps(iz1,jz0);
 259             dx11             = _mm_sub_ps(ix1,jx1);
 260             dy11             = _mm_sub_ps(iy1,jy1);
 261             dz11             = _mm_sub_ps(iz1,jz1);
 262             dx12             = _mm_sub_ps(ix1,jx2);
 263             dy12             = _mm_sub_ps(iy1,jy2);
 264             dz12             = _mm_sub_ps(iz1,jz2);
 265             dx20             = _mm_sub_ps(ix2,jx0);
 266             dy20             = _mm_sub_ps(iy2,jy0);
 267             dz20             = _mm_sub_ps(iz2,jz0);
 268             dx21             = _mm_sub_ps(ix2,jx1);
 269             dy21             = _mm_sub_ps(iy2,jy1);
 270             dz21             = _mm_sub_ps(iz2,jz1);
 271             dx22             = _mm_sub_ps(ix2,jx2);
 272             dy22             = _mm_sub_ps(iy2,jy2);
 273             dz22             = _mm_sub_ps(iz2,jz2);
 274
 275             /* Calculate squared distance and things based on it */
 276             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 277             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 278             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 279             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 280             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 281             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 282             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 283             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 284             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 285
 286             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 287             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 288             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 289             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 290             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 291             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 292             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 293             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 294             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 295
 296             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 297             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 298             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 299             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 300             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 301             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 302             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 303             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 304             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 305
 306             fjx0             = _mm_setzero_ps();
 307             fjy0             = _mm_setzero_ps();
 308             fjz0             = _mm_setzero_ps();
 309             fjx1             = _mm_setzero_ps();
 310             fjy1             = _mm_setzero_ps();
 311             fjz1             = _mm_setzero_ps();
 312             fjx2             = _mm_setzero_ps();
 313             fjy2             = _mm_setzero_ps();
 314             fjz2             = _mm_setzero_ps();
 315
 316             /**************************
 317              * CALCULATE INTERACTIONS *
 318              **************************/
 319
 320             if (gmx_mm_any_lt(rsq00,rcutoff2))
 321             {
 322
 323             r00              = _mm_mul_ps(rsq00,rinv00);
 324
 325             /* REACTION-FIELD ELECTROSTATICS */
 326             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 327             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 328
 329             /* LENNARD-JONES DISPERSION/REPULSION */
 330
 331             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 332             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 333             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 334             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 335             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 336
 337             d                = _mm_sub_ps(r00,rswitch);
 338             d                = _mm_max_ps(d,_mm_setzero_ps());
 339             d2               = _mm_mul_ps(d,d);
 340             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 341
 342             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 343
 344             /* Evaluate switch function */
 345             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 346             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 347             vvdw             = _mm_mul_ps(vvdw,sw);
 348             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 349
 350             /* Update potential sum for this i atom from the interaction with this j atom. */
 351             velec            = _mm_and_ps(velec,cutoff_mask);
 352             velecsum         = _mm_add_ps(velecsum,velec);
 353             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 354             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 355
 356             fscal            = _mm_add_ps(felec,fvdw);
 357
 358             fscal            = _mm_and_ps(fscal,cutoff_mask);
 359
 360              /* Update vectorial force */
 361             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 362             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 363             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 364
 365             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 366             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 367             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 368
 369             }
 370
 371             /**************************
 372              * CALCULATE INTERACTIONS *
 373              **************************/
 374
 375             if (gmx_mm_any_lt(rsq01,rcutoff2))
 376             {
 377
 378             /* REACTION-FIELD ELECTROSTATICS */
 379             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_macc_ps(krf,rsq01,rinv01),crf));
 380             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
 381
 382             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 383
 384             /* Update potential sum for this i atom from the interaction with this j atom. */
 385             velec            = _mm_and_ps(velec,cutoff_mask);
 386             velecsum         = _mm_add_ps(velecsum,velec);
 387
 388             fscal            = felec;
 389
 390             fscal            = _mm_and_ps(fscal,cutoff_mask);
 391
 392              /* Update vectorial force */
 393             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 394             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 395             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 396
 397             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 398             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 399             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 400
 401             }
 402
 403             /**************************
 404              * CALCULATE INTERACTIONS *
 405              **************************/
 406
 407             if (gmx_mm_any_lt(rsq02,rcutoff2))
 408             {
 409
 410             /* REACTION-FIELD ELECTROSTATICS */
 411             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_macc_ps(krf,rsq02,rinv02),crf));
 412             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
 413
 414             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 415
 416             /* Update potential sum for this i atom from the interaction with this j atom. */
 417             velec            = _mm_and_ps(velec,cutoff_mask);
 418             velecsum         = _mm_add_ps(velecsum,velec);
 419
 420             fscal            = felec;
 421
 422             fscal            = _mm_and_ps(fscal,cutoff_mask);
 423
 424              /* Update vectorial force */
 425             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 426             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 427             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 428
 429             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 430             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 431             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 432
 433             }
 434
 435             /**************************
 436              * CALCULATE INTERACTIONS *
 437              **************************/
 438
 439             if (gmx_mm_any_lt(rsq10,rcutoff2))
 440             {
 441
 442             /* REACTION-FIELD ELECTROSTATICS */
 443             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_macc_ps(krf,rsq10,rinv10),crf));
 444             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
 445
 446             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 447
 448             /* Update potential sum for this i atom from the interaction with this j atom. */
 449             velec            = _mm_and_ps(velec,cutoff_mask);
 450             velecsum         = _mm_add_ps(velecsum,velec);
 451
 452             fscal            = felec;
 453
 454             fscal            = _mm_and_ps(fscal,cutoff_mask);
 455
 456              /* Update vectorial force */
 457             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 458             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 459             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 460
 461             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 462             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 463             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 464
 465             }
 466
 467             /**************************
 468              * CALCULATE INTERACTIONS *
 469              **************************/
 470
 471             if (gmx_mm_any_lt(rsq11,rcutoff2))
 472             {
 473
 474             /* REACTION-FIELD ELECTROSTATICS */
 475             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 476             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 477
 478             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 479
 480             /* Update potential sum for this i atom from the interaction with this j atom. */
 481             velec            = _mm_and_ps(velec,cutoff_mask);
 482             velecsum         = _mm_add_ps(velecsum,velec);
 483
 484             fscal            = felec;
 485
 486             fscal            = _mm_and_ps(fscal,cutoff_mask);
 487
 488              /* Update vectorial force */
 489             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 490             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 491             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 492
 493             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 494             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 495             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 496
 497             }
 498
 499             /**************************
 500              * CALCULATE INTERACTIONS *
 501              **************************/
 502
 503             if (gmx_mm_any_lt(rsq12,rcutoff2))
 504             {
 505
 506             /* REACTION-FIELD ELECTROSTATICS */
 507             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 508             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 509
 510             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 511
 512             /* Update potential sum for this i atom from the interaction with this j atom. */
 513             velec            = _mm_and_ps(velec,cutoff_mask);
 514             velecsum         = _mm_add_ps(velecsum,velec);
 515
 516             fscal            = felec;
 517
 518             fscal            = _mm_and_ps(fscal,cutoff_mask);
 519
 520              /* Update vectorial force */
 521             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 522             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 523             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 524
 525             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 526             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 527             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 528
 529             }
 530
 531             /**************************
 532              * CALCULATE INTERACTIONS *
 533              **************************/
 534
 535             if (gmx_mm_any_lt(rsq20,rcutoff2))
 536             {
 537
 538             /* REACTION-FIELD ELECTROSTATICS */
 539             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_macc_ps(krf,rsq20,rinv20),crf));
 540             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
 541
 542             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 543
 544             /* Update potential sum for this i atom from the interaction with this j atom. */
 545             velec            = _mm_and_ps(velec,cutoff_mask);
 546             velecsum         = _mm_add_ps(velecsum,velec);
 547
 548             fscal            = felec;
 549
 550             fscal            = _mm_and_ps(fscal,cutoff_mask);
 551
 552              /* Update vectorial force */
 553             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 554             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 555             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 556
 557             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
 558             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
 559             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
 560
 561             }
 562
 563             /**************************
 564              * CALCULATE INTERACTIONS *
 565              **************************/
 566
 567             if (gmx_mm_any_lt(rsq21,rcutoff2))
 568             {
 569
 570             /* REACTION-FIELD ELECTROSTATICS */
 571             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 572             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 573
 574             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 575
 576             /* Update potential sum for this i atom from the interaction with this j atom. */
 577             velec            = _mm_and_ps(velec,cutoff_mask);
 578             velecsum         = _mm_add_ps(velecsum,velec);
 579
 580             fscal            = felec;
 581
 582             fscal            = _mm_and_ps(fscal,cutoff_mask);
 583
 584              /* Update vectorial force */
 585             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 586             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 587             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 588
 589             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 590             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 591             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 592
 593             }
 594
 595             /**************************
 596              * CALCULATE INTERACTIONS *
 597              **************************/
 598
 599             if (gmx_mm_any_lt(rsq22,rcutoff2))
 600             {
 601
 602             /* REACTION-FIELD ELECTROSTATICS */
 603             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 604             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 605
 606             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 607
 608             /* Update potential sum for this i atom from the interaction with this j atom. */
 609             velec            = _mm_and_ps(velec,cutoff_mask);
 610             velecsum         = _mm_add_ps(velecsum,velec);
 611
 612             fscal            = felec;
 613
 614             fscal            = _mm_and_ps(fscal,cutoff_mask);
 615
 616              /* Update vectorial force */
 617             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 618             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 619             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 620
 621             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 622             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 623             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 624
 625             }
 626
 627             fjptrA             = f+j_coord_offsetA;
 628             fjptrB             = f+j_coord_offsetB;
 629             fjptrC             = f+j_coord_offsetC;
 630             fjptrD             = f+j_coord_offsetD;
 631
 632             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 633                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 634
 635             /* Inner loop uses 385 flops */
 636         }
 637
 638         if(jidx<j_index_end)
 639         {
 640
 641             /* Get j neighbor index, and coordinate index */
 642             jnrlistA         = jjnr[jidx];
 643             jnrlistB         = jjnr[jidx+1];
 644             jnrlistC         = jjnr[jidx+2];
 645             jnrlistD         = jjnr[jidx+3];
 646             /* Sign of each element will be negative for non-real atoms.
 647              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 648              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 649              */
 650             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 651             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 652             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 653             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 654             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 655             j_coord_offsetA  = DIM*jnrA;
 656             j_coord_offsetB  = DIM*jnrB;
 657             j_coord_offsetC  = DIM*jnrC;
 658             j_coord_offsetD  = DIM*jnrD;
 659
 660             /* load j atom coordinates */
 661             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 662                                               x+j_coord_offsetC,x+j_coord_offsetD,
 663                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 664
 665             /* Calculate displacement vector */
 666             dx00             = _mm_sub_ps(ix0,jx0);
 667             dy00             = _mm_sub_ps(iy0,jy0);
 668             dz00             = _mm_sub_ps(iz0,jz0);
 669             dx01             = _mm_sub_ps(ix0,jx1);
 670             dy01             = _mm_sub_ps(iy0,jy1);
 671             dz01             = _mm_sub_ps(iz0,jz1);
 672             dx02             = _mm_sub_ps(ix0,jx2);
 673             dy02             = _mm_sub_ps(iy0,jy2);
 674             dz02             = _mm_sub_ps(iz0,jz2);
 675             dx10             = _mm_sub_ps(ix1,jx0);
 676             dy10             = _mm_sub_ps(iy1,jy0);
 677             dz10             = _mm_sub_ps(iz1,jz0);
 678             dx11             = _mm_sub_ps(ix1,jx1);
 679             dy11             = _mm_sub_ps(iy1,jy1);
 680             dz11             = _mm_sub_ps(iz1,jz1);
 681             dx12             = _mm_sub_ps(ix1,jx2);
 682             dy12             = _mm_sub_ps(iy1,jy2);
 683             dz12             = _mm_sub_ps(iz1,jz2);
 684             dx20             = _mm_sub_ps(ix2,jx0);
 685             dy20             = _mm_sub_ps(iy2,jy0);
 686             dz20             = _mm_sub_ps(iz2,jz0);
 687             dx21             = _mm_sub_ps(ix2,jx1);
 688             dy21             = _mm_sub_ps(iy2,jy1);
 689             dz21             = _mm_sub_ps(iz2,jz1);
 690             dx22             = _mm_sub_ps(ix2,jx2);
 691             dy22             = _mm_sub_ps(iy2,jy2);
 692             dz22             = _mm_sub_ps(iz2,jz2);
 693
 694             /* Calculate squared distance and things based on it */
 695             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 696             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
 697             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
 698             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
 699             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 700             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 701             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
 702             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 703             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 704
 705             rinv00           = gmx_mm_invsqrt_ps(rsq00);
 706             rinv01           = gmx_mm_invsqrt_ps(rsq01);
 707             rinv02           = gmx_mm_invsqrt_ps(rsq02);
 708             rinv10           = gmx_mm_invsqrt_ps(rsq10);
 709             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 710             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 711             rinv20           = gmx_mm_invsqrt_ps(rsq20);
 712             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 713             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 714
 715             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
 716             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
 717             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
 718             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
 719             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 720             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 721             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
 722             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 723             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 724
 725             fjx0             = _mm_setzero_ps();
 726             fjy0             = _mm_setzero_ps();
 727             fjz0             = _mm_setzero_ps();
 728             fjx1             = _mm_setzero_ps();
 729             fjy1             = _mm_setzero_ps();
 730             fjz1             = _mm_setzero_ps();
 731             fjx2             = _mm_setzero_ps();
 732             fjy2             = _mm_setzero_ps();
 733             fjz2             = _mm_setzero_ps();
 734
 735             /**************************
 736              * CALCULATE INTERACTIONS *
 737              **************************/
 738
 739             if (gmx_mm_any_lt(rsq00,rcutoff2))
 740             {
 741
 742             r00              = _mm_mul_ps(rsq00,rinv00);
 743             r00              = _mm_andnot_ps(dummy_mask,r00);
 744
 745             /* REACTION-FIELD ELECTROSTATICS */
 746             velec            = _mm_mul_ps(qq00,_mm_sub_ps(_mm_macc_ps(krf,rsq00,rinv00),crf));
 747             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
 748
 749             /* LENNARD-JONES DISPERSION/REPULSION */
 750
 751             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 752             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 753             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 754             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 755             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 756
 757             d                = _mm_sub_ps(r00,rswitch);
 758             d                = _mm_max_ps(d,_mm_setzero_ps());
 759             d2               = _mm_mul_ps(d,d);
 760             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
 761
 762             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
 763
 764             /* Evaluate switch function */
 765             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
 766             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
 767             vvdw             = _mm_mul_ps(vvdw,sw);
 768             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 769
 770             /* Update potential sum for this i atom from the interaction with this j atom. */
 771             velec            = _mm_and_ps(velec,cutoff_mask);
 772             velec            = _mm_andnot_ps(dummy_mask,velec);
 773             velecsum         = _mm_add_ps(velecsum,velec);
 774             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 775             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 776             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 777
 778             fscal            = _mm_add_ps(felec,fvdw);
 779
 780             fscal            = _mm_and_ps(fscal,cutoff_mask);
 781
 782             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 783
 784              /* Update vectorial force */
 785             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 786             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 787             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 788
 789             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 790             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 791             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 792
 793             }
 794
 795             /**************************
 796              * CALCULATE INTERACTIONS *
 797              **************************/
 798
 799             if (gmx_mm_any_lt(rsq01,rcutoff2))
 800             {
 801
 802             /* REACTION-FIELD ELECTROSTATICS */
 803             velec            = _mm_mul_ps(qq01,_mm_sub_ps(_mm_macc_ps(krf,rsq01,rinv01),crf));
 804             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
 805
 806             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
 807
 808             /* Update potential sum for this i atom from the interaction with this j atom. */
 809             velec            = _mm_and_ps(velec,cutoff_mask);
 810             velec            = _mm_andnot_ps(dummy_mask,velec);
 811             velecsum         = _mm_add_ps(velecsum,velec);
 812
 813             fscal            = felec;
 814
 815             fscal            = _mm_and_ps(fscal,cutoff_mask);
 816
 817             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 818
 819              /* Update vectorial force */
 820             fix0             = _mm_macc_ps(dx01,fscal,fix0);
 821             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
 822             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
 823
 824             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
 825             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
 826             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
 827
 828             }
 829
 830             /**************************
 831              * CALCULATE INTERACTIONS *
 832              **************************/
 833
 834             if (gmx_mm_any_lt(rsq02,rcutoff2))
 835             {
 836
 837             /* REACTION-FIELD ELECTROSTATICS */
 838             velec            = _mm_mul_ps(qq02,_mm_sub_ps(_mm_macc_ps(krf,rsq02,rinv02),crf));
 839             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
 840
 841             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
 842
 843             /* Update potential sum for this i atom from the interaction with this j atom. */
 844             velec            = _mm_and_ps(velec,cutoff_mask);
 845             velec            = _mm_andnot_ps(dummy_mask,velec);
 846             velecsum         = _mm_add_ps(velecsum,velec);
 847
 848             fscal            = felec;
 849
 850             fscal            = _mm_and_ps(fscal,cutoff_mask);
 851
 852             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 853
 854              /* Update vectorial force */
 855             fix0             = _mm_macc_ps(dx02,fscal,fix0);
 856             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
 857             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
 858
 859             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
 860             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
 861             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
 862
 863             }
 864
 865             /**************************
 866              * CALCULATE INTERACTIONS *
 867              **************************/
 868
 869             if (gmx_mm_any_lt(rsq10,rcutoff2))
 870             {
 871
 872             /* REACTION-FIELD ELECTROSTATICS */
 873             velec            = _mm_mul_ps(qq10,_mm_sub_ps(_mm_macc_ps(krf,rsq10,rinv10),crf));
 874             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
 875
 876             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
 877
 878             /* Update potential sum for this i atom from the interaction with this j atom. */
 879             velec            = _mm_and_ps(velec,cutoff_mask);
 880             velec            = _mm_andnot_ps(dummy_mask,velec);
 881             velecsum         = _mm_add_ps(velecsum,velec);
 882
 883             fscal            = felec;
 884
 885             fscal            = _mm_and_ps(fscal,cutoff_mask);
 886
 887             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 888
 889              /* Update vectorial force */
 890             fix1             = _mm_macc_ps(dx10,fscal,fix1);
 891             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
 892             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
 893
 894             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
 895             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
 896             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
 897
 898             }
 899
 900             /**************************
 901              * CALCULATE INTERACTIONS *
 902              **************************/
 903
 904             if (gmx_mm_any_lt(rsq11,rcutoff2))
 905             {
 906
 907             /* REACTION-FIELD ELECTROSTATICS */
 908             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 909             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 910
 911             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 912
 913             /* Update potential sum for this i atom from the interaction with this j atom. */
 914             velec            = _mm_and_ps(velec,cutoff_mask);
 915             velec            = _mm_andnot_ps(dummy_mask,velec);
 916             velecsum         = _mm_add_ps(velecsum,velec);
 917
 918             fscal            = felec;
 919
 920             fscal            = _mm_and_ps(fscal,cutoff_mask);
 921
 922             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 923
 924              /* Update vectorial force */
 925             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 926             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 927             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 928
 929             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 930             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 931             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 932
 933             }
 934
 935             /**************************
 936              * CALCULATE INTERACTIONS *
 937              **************************/
 938
 939             if (gmx_mm_any_lt(rsq12,rcutoff2))
 940             {
 941
 942             /* REACTION-FIELD ELECTROSTATICS */
 943             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 944             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 945
 946             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 947
 948             /* Update potential sum for this i atom from the interaction with this j atom. */
 949             velec            = _mm_and_ps(velec,cutoff_mask);
 950             velec            = _mm_andnot_ps(dummy_mask,velec);
 951             velecsum         = _mm_add_ps(velecsum,velec);
 952
 953             fscal            = felec;
 954
 955             fscal            = _mm_and_ps(fscal,cutoff_mask);
 956
 957             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 958
 959              /* Update vectorial force */
 960             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 961             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 962             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 963
 964             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 965             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 966             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 967
 968             }
 969
 970             /**************************
 971              * CALCULATE INTERACTIONS *
 972              **************************/
 973
 974             if (gmx_mm_any_lt(rsq20,rcutoff2))
 975             {
 976
 977             /* REACTION-FIELD ELECTROSTATICS */
 978             velec            = _mm_mul_ps(qq20,_mm_sub_ps(_mm_macc_ps(krf,rsq20,rinv20),crf));
 979             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
 980
 981             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
 982
 983             /* Update potential sum for this i atom from the interaction with this j atom. */
 984             velec            = _mm_and_ps(velec,cutoff_mask);
 985             velec            = _mm_andnot_ps(dummy_mask,velec);
 986             velecsum         = _mm_add_ps(velecsum,velec);
 987
 988             fscal            = felec;
 989
 990             fscal            = _mm_and_ps(fscal,cutoff_mask);
 991
 992             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 993
 994              /* Update vectorial force */
 995             fix2             = _mm_macc_ps(dx20,fscal,fix2);
 996             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
 997             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
 998
 999             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1000             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1001             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1002
1003             }
1004
1005             /**************************
1006              * CALCULATE INTERACTIONS *
1007              **************************/
1008
1009             if (gmx_mm_any_lt(rsq21,rcutoff2))
1010             {
1011
1012             /* REACTION-FIELD ELECTROSTATICS */
1013             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
1014             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1015
1016             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1017
1018             /* Update potential sum for this i atom from the interaction with this j atom. */
1019             velec            = _mm_and_ps(velec,cutoff_mask);
1020             velec            = _mm_andnot_ps(dummy_mask,velec);
1021             velecsum         = _mm_add_ps(velecsum,velec);
1022
1023             fscal            = felec;
1024
1025             fscal            = _mm_and_ps(fscal,cutoff_mask);
1026
1027             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1028
1029              /* Update vectorial force */
1030             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1031             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1032             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1033
1034             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1035             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1036             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1037
1038             }
1039
1040             /**************************
1041              * CALCULATE INTERACTIONS *
1042              **************************/
1043
1044             if (gmx_mm_any_lt(rsq22,rcutoff2))
1045             {
1046
1047             /* REACTION-FIELD ELECTROSTATICS */
1048             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
1049             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1050
1051             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1052
1053             /* Update potential sum for this i atom from the interaction with this j atom. */
1054             velec            = _mm_and_ps(velec,cutoff_mask);
1055             velec            = _mm_andnot_ps(dummy_mask,velec);
1056             velecsum         = _mm_add_ps(velecsum,velec);
1057
1058             fscal            = felec;
1059
1060             fscal            = _mm_and_ps(fscal,cutoff_mask);
1061
1062             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1063
1064              /* Update vectorial force */
1065             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1066             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1067             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1068
1069             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1070             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1071             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1072
1073             }
1074
1075             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1076             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1077             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1078             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1079
1080             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1081                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1082
1083             /* Inner loop uses 386 flops */
1084         }
1085
1086         /* End of innermost loop */
1087
1088         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1089                                               f+i_coord_offset,fshift+i_shift_offset);
1090
1091         ggid                        = gid[iidx];
1092         /* Update potential energies */
1093         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1094         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1095
1096         /* Increment number of inner iterations */
1097         inneriter                  += j_index_end - j_index_start;
1098
1099         /* Outer loop uses 20 flops */
1100     }
1101
1102     /* Increment number of outer iterations */
1103     outeriter        += nri;
1104
1105     /* Update outer/inner flops */
1106
1107     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*386);
1108 }
1109 /*
1110  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSw_GeomW3W3_F_avx_128_fma_single
1111  * Electrostatics interaction: ReactionField
1112  * VdW interaction:            LennardJones
1113  * Geometry:                   Water3-Water3
1114  * Calculate force/pot:        Force
1115  */
1116 void
1117 nb_kernel_ElecRFCut_VdwLJSw_GeomW3W3_F_avx_128_fma_single
1118                     (t_nblist                    * gmx_restrict       nlist,
1119                      rvec                        * gmx_restrict          xx,
1120                      rvec                        * gmx_restrict          ff,
1121                      t_forcerec                  * gmx_restrict          fr,
1122                      t_mdatoms                   * gmx_restrict     mdatoms,
1123                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1124                      t_nrnb                      * gmx_restrict        nrnb)
1125 {
1126     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1127      * just 0 for non-waters.
1128      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1129      * jnr indices corresponding to data put in the four positions in the SIMD register.
1130      */
1131     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1132     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1133     int              jnrA,jnrB,jnrC,jnrD;
1134     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1135     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1136     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1137     real             rcutoff_scalar;
1138     real             *shiftvec,*fshift,*x,*f;
1139     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1140     real             scratch[4*DIM];
1141     __m128           fscal,rcutoff,rcutoff2,jidxall;
1142     int              vdwioffset0;
1143     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1144     int              vdwioffset1;
1145     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1146     int              vdwioffset2;
1147     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1148     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1149     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1150     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1151     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1152     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1153     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1154     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1155     __m128           dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1156     __m128           dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1157     __m128           dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1158     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1159     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1160     __m128           dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1161     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1162     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1163     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1164     real             *charge;
1165     int              nvdwtype;
1166     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1167     int              *vdwtype;
1168     real             *vdwparam;
1169     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1170     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1171     __m128           rswitch,swV3,swV4,swV5,swF2,swF3,swF4,d,d2,sw,dsw;
1172     real             rswitch_scalar,d_scalar;
1173     __m128           dummy_mask,cutoff_mask;
1174     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1175     __m128           one     = _mm_set1_ps(1.0);
1176     __m128           two     = _mm_set1_ps(2.0);
1177     x                = xx[0];
1178     f                = ff[0];
1179
1180     nri              = nlist->nri;
1181     iinr             = nlist->iinr;
1182     jindex           = nlist->jindex;
1183     jjnr             = nlist->jjnr;
1184     shiftidx         = nlist->shift;
1185     gid              = nlist->gid;
1186     shiftvec         = fr->shift_vec[0];
1187     fshift           = fr->fshift[0];
1188     facel            = _mm_set1_ps(fr->epsfac);
1189     charge           = mdatoms->chargeA;
1190     krf              = _mm_set1_ps(fr->ic->k_rf);
1191     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1192     crf              = _mm_set1_ps(fr->ic->c_rf);
1193     nvdwtype         = fr->ntype;
1194     vdwparam         = fr->nbfp;
1195     vdwtype          = mdatoms->typeA;
1196
1197     /* Setup water-specific parameters */
1198     inr              = nlist->iinr[0];
1199     iq0              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+0]));
1200     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1201     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1202     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1203
1204     jq0              = _mm_set1_ps(charge[inr+0]);
1205     jq1              = _mm_set1_ps(charge[inr+1]);
1206     jq2              = _mm_set1_ps(charge[inr+2]);
1207     vdwjidx0A        = 2*vdwtype[inr+0];
1208     qq00             = _mm_mul_ps(iq0,jq0);
1209     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1210     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1211     qq01             = _mm_mul_ps(iq0,jq1);
1212     qq02             = _mm_mul_ps(iq0,jq2);
1213     qq10             = _mm_mul_ps(iq1,jq0);
1214     qq11             = _mm_mul_ps(iq1,jq1);
1215     qq12             = _mm_mul_ps(iq1,jq2);
1216     qq20             = _mm_mul_ps(iq2,jq0);
1217     qq21             = _mm_mul_ps(iq2,jq1);
1218     qq22             = _mm_mul_ps(iq2,jq2);
1219
1220     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1221     rcutoff_scalar   = fr->rcoulomb;
1222     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1223     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1224
1225     rswitch_scalar   = fr->rvdw_switch;
1226     rswitch          = _mm_set1_ps(rswitch_scalar);
1227     /* Setup switch parameters */
1228     d_scalar         = rcutoff_scalar-rswitch_scalar;
1229     d                = _mm_set1_ps(d_scalar);
1230     swV3             = _mm_set1_ps(-10.0/(d_scalar*d_scalar*d_scalar));
1231     swV4             = _mm_set1_ps( 15.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1232     swV5             = _mm_set1_ps( -6.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1233     swF2             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar));
1234     swF3             = _mm_set1_ps( 60.0/(d_scalar*d_scalar*d_scalar*d_scalar));
1235     swF4             = _mm_set1_ps(-30.0/(d_scalar*d_scalar*d_scalar*d_scalar*d_scalar));
1236
1237     /* Avoid stupid compiler warnings */
1238     jnrA = jnrB = jnrC = jnrD = 0;
1239     j_coord_offsetA = 0;
1240     j_coord_offsetB = 0;
1241     j_coord_offsetC = 0;
1242     j_coord_offsetD = 0;
1243
1244     outeriter        = 0;
1245     inneriter        = 0;
1246
1247     for(iidx=0;iidx<4*DIM;iidx++)
1248     {
1249         scratch[iidx] = 0.0;
1250     }
1251
1252     /* Start outer loop over neighborlists */
1253     for(iidx=0; iidx<nri; iidx++)
1254     {
1255         /* Load shift vector for this list */
1256         i_shift_offset   = DIM*shiftidx[iidx];
1257
1258         /* Load limits for loop over neighbors */
1259         j_index_start    = jindex[iidx];
1260         j_index_end      = jindex[iidx+1];
1261
1262         /* Get outer coordinate index */
1263         inr              = iinr[iidx];
1264         i_coord_offset   = DIM*inr;
1265
1266         /* Load i particle coords and add shift vector */
1267         gmx_mm_load_shift_and_3rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1268                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1269
1270         fix0             = _mm_setzero_ps();
1271         fiy0             = _mm_setzero_ps();
1272         fiz0             = _mm_setzero_ps();
1273         fix1             = _mm_setzero_ps();
1274         fiy1             = _mm_setzero_ps();
1275         fiz1             = _mm_setzero_ps();
1276         fix2             = _mm_setzero_ps();
1277         fiy2             = _mm_setzero_ps();
1278         fiz2             = _mm_setzero_ps();
1279
1280         /* Start inner kernel loop */
1281         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1282         {
1283
1284             /* Get j neighbor index, and coordinate index */
1285             jnrA             = jjnr[jidx];
1286             jnrB             = jjnr[jidx+1];
1287             jnrC             = jjnr[jidx+2];
1288             jnrD             = jjnr[jidx+3];
1289             j_coord_offsetA  = DIM*jnrA;
1290             j_coord_offsetB  = DIM*jnrB;
1291             j_coord_offsetC  = DIM*jnrC;
1292             j_coord_offsetD  = DIM*jnrD;
1293
1294             /* load j atom coordinates */
1295             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1296                                               x+j_coord_offsetC,x+j_coord_offsetD,
1297                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1298
1299             /* Calculate displacement vector */
1300             dx00             = _mm_sub_ps(ix0,jx0);
1301             dy00             = _mm_sub_ps(iy0,jy0);
1302             dz00             = _mm_sub_ps(iz0,jz0);
1303             dx01             = _mm_sub_ps(ix0,jx1);
1304             dy01             = _mm_sub_ps(iy0,jy1);
1305             dz01             = _mm_sub_ps(iz0,jz1);
1306             dx02             = _mm_sub_ps(ix0,jx2);
1307             dy02             = _mm_sub_ps(iy0,jy2);
1308             dz02             = _mm_sub_ps(iz0,jz2);
1309             dx10             = _mm_sub_ps(ix1,jx0);
1310             dy10             = _mm_sub_ps(iy1,jy0);
1311             dz10             = _mm_sub_ps(iz1,jz0);
1312             dx11             = _mm_sub_ps(ix1,jx1);
1313             dy11             = _mm_sub_ps(iy1,jy1);
1314             dz11             = _mm_sub_ps(iz1,jz1);
1315             dx12             = _mm_sub_ps(ix1,jx2);
1316             dy12             = _mm_sub_ps(iy1,jy2);
1317             dz12             = _mm_sub_ps(iz1,jz2);
1318             dx20             = _mm_sub_ps(ix2,jx0);
1319             dy20             = _mm_sub_ps(iy2,jy0);
1320             dz20             = _mm_sub_ps(iz2,jz0);
1321             dx21             = _mm_sub_ps(ix2,jx1);
1322             dy21             = _mm_sub_ps(iy2,jy1);
1323             dz21             = _mm_sub_ps(iz2,jz1);
1324             dx22             = _mm_sub_ps(ix2,jx2);
1325             dy22             = _mm_sub_ps(iy2,jy2);
1326             dz22             = _mm_sub_ps(iz2,jz2);
1327
1328             /* Calculate squared distance and things based on it */
1329             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1330             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1331             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1332             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1333             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1334             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1335             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1336             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1337             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1338
1339             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1340             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1341             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1342             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1343             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1344             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1345             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1346             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1347             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1348
1349             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1350             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1351             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1352             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1353             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1354             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1355             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1356             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1357             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1358
1359             fjx0             = _mm_setzero_ps();
1360             fjy0             = _mm_setzero_ps();
1361             fjz0             = _mm_setzero_ps();
1362             fjx1             = _mm_setzero_ps();
1363             fjy1             = _mm_setzero_ps();
1364             fjz1             = _mm_setzero_ps();
1365             fjx2             = _mm_setzero_ps();
1366             fjy2             = _mm_setzero_ps();
1367             fjz2             = _mm_setzero_ps();
1368
1369             /**************************
1370              * CALCULATE INTERACTIONS *
1371              **************************/
1372
1373             if (gmx_mm_any_lt(rsq00,rcutoff2))
1374             {
1375
1376             r00              = _mm_mul_ps(rsq00,rinv00);
1377
1378             /* REACTION-FIELD ELECTROSTATICS */
1379             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
1380
1381             /* LENNARD-JONES DISPERSION/REPULSION */
1382
1383             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1384             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
1385             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
1386             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
1387             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
1388
1389             d                = _mm_sub_ps(r00,rswitch);
1390             d                = _mm_max_ps(d,_mm_setzero_ps());
1391             d2               = _mm_mul_ps(d,d);
1392             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
1393
1394             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
1395
1396             /* Evaluate switch function */
1397             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1398             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
1399             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1400
1401             fscal            = _mm_add_ps(felec,fvdw);
1402
1403             fscal            = _mm_and_ps(fscal,cutoff_mask);
1404
1405              /* Update vectorial force */
1406             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1407             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1408             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1409
1410             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1411             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1412             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1413
1414             }
1415
1416             /**************************
1417              * CALCULATE INTERACTIONS *
1418              **************************/
1419
1420             if (gmx_mm_any_lt(rsq01,rcutoff2))
1421             {
1422
1423             /* REACTION-FIELD ELECTROSTATICS */
1424             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
1425
1426             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1427
1428             fscal            = felec;
1429
1430             fscal            = _mm_and_ps(fscal,cutoff_mask);
1431
1432              /* Update vectorial force */
1433             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1434             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1435             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1436
1437             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1438             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1439             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1440
1441             }
1442
1443             /**************************
1444              * CALCULATE INTERACTIONS *
1445              **************************/
1446
1447             if (gmx_mm_any_lt(rsq02,rcutoff2))
1448             {
1449
1450             /* REACTION-FIELD ELECTROSTATICS */
1451             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
1452
1453             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
1454
1455             fscal            = felec;
1456
1457             fscal            = _mm_and_ps(fscal,cutoff_mask);
1458
1459              /* Update vectorial force */
1460             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1461             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1462             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1463
1464             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1465             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1466             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1467
1468             }
1469
1470             /**************************
1471              * CALCULATE INTERACTIONS *
1472              **************************/
1473
1474             if (gmx_mm_any_lt(rsq10,rcutoff2))
1475             {
1476
1477             /* REACTION-FIELD ELECTROSTATICS */
1478             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
1479
1480             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1481
1482             fscal            = felec;
1483
1484             fscal            = _mm_and_ps(fscal,cutoff_mask);
1485
1486              /* Update vectorial force */
1487             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1488             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1489             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1490
1491             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1492             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1493             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1494
1495             }
1496
1497             /**************************
1498              * CALCULATE INTERACTIONS *
1499              **************************/
1500
1501             if (gmx_mm_any_lt(rsq11,rcutoff2))
1502             {
1503
1504             /* REACTION-FIELD ELECTROSTATICS */
1505             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1506
1507             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1508
1509             fscal            = felec;
1510
1511             fscal            = _mm_and_ps(fscal,cutoff_mask);
1512
1513              /* Update vectorial force */
1514             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1515             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1516             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1517
1518             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1519             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1520             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1521
1522             }
1523
1524             /**************************
1525              * CALCULATE INTERACTIONS *
1526              **************************/
1527
1528             if (gmx_mm_any_lt(rsq12,rcutoff2))
1529             {
1530
1531             /* REACTION-FIELD ELECTROSTATICS */
1532             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1533
1534             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1535
1536             fscal            = felec;
1537
1538             fscal            = _mm_and_ps(fscal,cutoff_mask);
1539
1540              /* Update vectorial force */
1541             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1542             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1543             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1544
1545             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1546             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1547             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1548
1549             }
1550
1551             /**************************
1552              * CALCULATE INTERACTIONS *
1553              **************************/
1554
1555             if (gmx_mm_any_lt(rsq20,rcutoff2))
1556             {
1557
1558             /* REACTION-FIELD ELECTROSTATICS */
1559             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
1560
1561             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1562
1563             fscal            = felec;
1564
1565             fscal            = _mm_and_ps(fscal,cutoff_mask);
1566
1567              /* Update vectorial force */
1568             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1569             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1570             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1571
1572             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1573             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1574             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1575
1576             }
1577
1578             /**************************
1579              * CALCULATE INTERACTIONS *
1580              **************************/
1581
1582             if (gmx_mm_any_lt(rsq21,rcutoff2))
1583             {
1584
1585             /* REACTION-FIELD ELECTROSTATICS */
1586             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1587
1588             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1589
1590             fscal            = felec;
1591
1592             fscal            = _mm_and_ps(fscal,cutoff_mask);
1593
1594              /* Update vectorial force */
1595             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1596             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1597             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1598
1599             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1600             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1601             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1602
1603             }
1604
1605             /**************************
1606              * CALCULATE INTERACTIONS *
1607              **************************/
1608
1609             if (gmx_mm_any_lt(rsq22,rcutoff2))
1610             {
1611
1612             /* REACTION-FIELD ELECTROSTATICS */
1613             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1614
1615             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1616
1617             fscal            = felec;
1618
1619             fscal            = _mm_and_ps(fscal,cutoff_mask);
1620
1621              /* Update vectorial force */
1622             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1623             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1624             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1625
1626             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1627             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1628             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1629
1630             }
1631
1632             fjptrA             = f+j_coord_offsetA;
1633             fjptrB             = f+j_coord_offsetB;
1634             fjptrC             = f+j_coord_offsetC;
1635             fjptrD             = f+j_coord_offsetD;
1636
1637             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1638                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1639
1640             /* Inner loop uses 328 flops */
1641         }
1642
1643         if(jidx<j_index_end)
1644         {
1645
1646             /* Get j neighbor index, and coordinate index */
1647             jnrlistA         = jjnr[jidx];
1648             jnrlistB         = jjnr[jidx+1];
1649             jnrlistC         = jjnr[jidx+2];
1650             jnrlistD         = jjnr[jidx+3];
1651             /* Sign of each element will be negative for non-real atoms.
1652              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1653              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1654              */
1655             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1656             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1657             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1658             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1659             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1660             j_coord_offsetA  = DIM*jnrA;
1661             j_coord_offsetB  = DIM*jnrB;
1662             j_coord_offsetC  = DIM*jnrC;
1663             j_coord_offsetD  = DIM*jnrD;
1664
1665             /* load j atom coordinates */
1666             gmx_mm_load_3rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1667                                               x+j_coord_offsetC,x+j_coord_offsetD,
1668                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1669
1670             /* Calculate displacement vector */
1671             dx00             = _mm_sub_ps(ix0,jx0);
1672             dy00             = _mm_sub_ps(iy0,jy0);
1673             dz00             = _mm_sub_ps(iz0,jz0);
1674             dx01             = _mm_sub_ps(ix0,jx1);
1675             dy01             = _mm_sub_ps(iy0,jy1);
1676             dz01             = _mm_sub_ps(iz0,jz1);
1677             dx02             = _mm_sub_ps(ix0,jx2);
1678             dy02             = _mm_sub_ps(iy0,jy2);
1679             dz02             = _mm_sub_ps(iz0,jz2);
1680             dx10             = _mm_sub_ps(ix1,jx0);
1681             dy10             = _mm_sub_ps(iy1,jy0);
1682             dz10             = _mm_sub_ps(iz1,jz0);
1683             dx11             = _mm_sub_ps(ix1,jx1);
1684             dy11             = _mm_sub_ps(iy1,jy1);
1685             dz11             = _mm_sub_ps(iz1,jz1);
1686             dx12             = _mm_sub_ps(ix1,jx2);
1687             dy12             = _mm_sub_ps(iy1,jy2);
1688             dz12             = _mm_sub_ps(iz1,jz2);
1689             dx20             = _mm_sub_ps(ix2,jx0);
1690             dy20             = _mm_sub_ps(iy2,jy0);
1691             dz20             = _mm_sub_ps(iz2,jz0);
1692             dx21             = _mm_sub_ps(ix2,jx1);
1693             dy21             = _mm_sub_ps(iy2,jy1);
1694             dz21             = _mm_sub_ps(iz2,jz1);
1695             dx22             = _mm_sub_ps(ix2,jx2);
1696             dy22             = _mm_sub_ps(iy2,jy2);
1697             dz22             = _mm_sub_ps(iz2,jz2);
1698
1699             /* Calculate squared distance and things based on it */
1700             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1701             rsq01            = gmx_mm_calc_rsq_ps(dx01,dy01,dz01);
1702             rsq02            = gmx_mm_calc_rsq_ps(dx02,dy02,dz02);
1703             rsq10            = gmx_mm_calc_rsq_ps(dx10,dy10,dz10);
1704             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1705             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1706             rsq20            = gmx_mm_calc_rsq_ps(dx20,dy20,dz20);
1707             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1708             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1709
1710             rinv00           = gmx_mm_invsqrt_ps(rsq00);
1711             rinv01           = gmx_mm_invsqrt_ps(rsq01);
1712             rinv02           = gmx_mm_invsqrt_ps(rsq02);
1713             rinv10           = gmx_mm_invsqrt_ps(rsq10);
1714             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1715             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1716             rinv20           = gmx_mm_invsqrt_ps(rsq20);
1717             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1718             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1719
1720             rinvsq00         = _mm_mul_ps(rinv00,rinv00);
1721             rinvsq01         = _mm_mul_ps(rinv01,rinv01);
1722             rinvsq02         = _mm_mul_ps(rinv02,rinv02);
1723             rinvsq10         = _mm_mul_ps(rinv10,rinv10);
1724             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1725             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1726             rinvsq20         = _mm_mul_ps(rinv20,rinv20);
1727             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1728             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1729
1730             fjx0             = _mm_setzero_ps();
1731             fjy0             = _mm_setzero_ps();
1732             fjz0             = _mm_setzero_ps();
1733             fjx1             = _mm_setzero_ps();
1734             fjy1             = _mm_setzero_ps();
1735             fjz1             = _mm_setzero_ps();
1736             fjx2             = _mm_setzero_ps();
1737             fjy2             = _mm_setzero_ps();
1738             fjz2             = _mm_setzero_ps();
1739
1740             /**************************
1741              * CALCULATE INTERACTIONS *
1742              **************************/
1743
1744             if (gmx_mm_any_lt(rsq00,rcutoff2))
1745             {
1746
1747             r00              = _mm_mul_ps(rsq00,rinv00);
1748             r00              = _mm_andnot_ps(dummy_mask,r00);
1749
1750             /* REACTION-FIELD ELECTROSTATICS */
1751             felec            = _mm_mul_ps(qq00,_mm_msub_ps(rinv00,rinvsq00,krf2));
1752
1753             /* LENNARD-JONES DISPERSION/REPULSION */
1754
1755             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1756             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
1757             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
1758             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
1759             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
1760
1761             d                = _mm_sub_ps(r00,rswitch);
1762             d                = _mm_max_ps(d,_mm_setzero_ps());
1763             d2               = _mm_mul_ps(d,d);
1764             sw               = _mm_add_ps(one,_mm_mul_ps(d2,_mm_mul_ps(d,_mm_macc_ps(d,_mm_macc_ps(d,swV5,swV4),swV3))));
1765
1766             dsw              = _mm_mul_ps(d2,_mm_macc_ps(d,_mm_macc_ps(d,swF4,swF3),swF2));
1767
1768             /* Evaluate switch function */
1769             /* fscal'=f'/r=-(v*sw)'/r=-(v'*sw+v*dsw)/r=-v'*sw/r-v*dsw/r=fscal*sw-v*dsw/r */
1770             fvdw             = _mm_msub_ps( fvdw,sw , _mm_mul_ps(rinv00,_mm_mul_ps(vvdw,dsw)) );
1771             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1772
1773             fscal            = _mm_add_ps(felec,fvdw);
1774
1775             fscal            = _mm_and_ps(fscal,cutoff_mask);
1776
1777             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1778
1779              /* Update vectorial force */
1780             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1781             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1782             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1783
1784             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1785             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1786             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1787
1788             }
1789
1790             /**************************
1791              * CALCULATE INTERACTIONS *
1792              **************************/
1793
1794             if (gmx_mm_any_lt(rsq01,rcutoff2))
1795             {
1796
1797             /* REACTION-FIELD ELECTROSTATICS */
1798             felec            = _mm_mul_ps(qq01,_mm_msub_ps(rinv01,rinvsq01,krf2));
1799
1800             cutoff_mask      = _mm_cmplt_ps(rsq01,rcutoff2);
1801
1802             fscal            = felec;
1803
1804             fscal            = _mm_and_ps(fscal,cutoff_mask);
1805
1806             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1807
1808              /* Update vectorial force */
1809             fix0             = _mm_macc_ps(dx01,fscal,fix0);
1810             fiy0             = _mm_macc_ps(dy01,fscal,fiy0);
1811             fiz0             = _mm_macc_ps(dz01,fscal,fiz0);
1812
1813             fjx1             = _mm_macc_ps(dx01,fscal,fjx1);
1814             fjy1             = _mm_macc_ps(dy01,fscal,fjy1);
1815             fjz1             = _mm_macc_ps(dz01,fscal,fjz1);
1816
1817             }
1818
1819             /**************************
1820              * CALCULATE INTERACTIONS *
1821              **************************/
1822
1823             if (gmx_mm_any_lt(rsq02,rcutoff2))
1824             {
1825
1826             /* REACTION-FIELD ELECTROSTATICS */
1827             felec            = _mm_mul_ps(qq02,_mm_msub_ps(rinv02,rinvsq02,krf2));
1828
1829             cutoff_mask      = _mm_cmplt_ps(rsq02,rcutoff2);
1830
1831             fscal            = felec;
1832
1833             fscal            = _mm_and_ps(fscal,cutoff_mask);
1834
1835             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1836
1837              /* Update vectorial force */
1838             fix0             = _mm_macc_ps(dx02,fscal,fix0);
1839             fiy0             = _mm_macc_ps(dy02,fscal,fiy0);
1840             fiz0             = _mm_macc_ps(dz02,fscal,fiz0);
1841
1842             fjx2             = _mm_macc_ps(dx02,fscal,fjx2);
1843             fjy2             = _mm_macc_ps(dy02,fscal,fjy2);
1844             fjz2             = _mm_macc_ps(dz02,fscal,fjz2);
1845
1846             }
1847
1848             /**************************
1849              * CALCULATE INTERACTIONS *
1850              **************************/
1851
1852             if (gmx_mm_any_lt(rsq10,rcutoff2))
1853             {
1854
1855             /* REACTION-FIELD ELECTROSTATICS */
1856             felec            = _mm_mul_ps(qq10,_mm_msub_ps(rinv10,rinvsq10,krf2));
1857
1858             cutoff_mask      = _mm_cmplt_ps(rsq10,rcutoff2);
1859
1860             fscal            = felec;
1861
1862             fscal            = _mm_and_ps(fscal,cutoff_mask);
1863
1864             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1865
1866              /* Update vectorial force */
1867             fix1             = _mm_macc_ps(dx10,fscal,fix1);
1868             fiy1             = _mm_macc_ps(dy10,fscal,fiy1);
1869             fiz1             = _mm_macc_ps(dz10,fscal,fiz1);
1870
1871             fjx0             = _mm_macc_ps(dx10,fscal,fjx0);
1872             fjy0             = _mm_macc_ps(dy10,fscal,fjy0);
1873             fjz0             = _mm_macc_ps(dz10,fscal,fjz0);
1874
1875             }
1876
1877             /**************************
1878              * CALCULATE INTERACTIONS *
1879              **************************/
1880
1881             if (gmx_mm_any_lt(rsq11,rcutoff2))
1882             {
1883
1884             /* REACTION-FIELD ELECTROSTATICS */
1885             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1886
1887             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1888
1889             fscal            = felec;
1890
1891             fscal            = _mm_and_ps(fscal,cutoff_mask);
1892
1893             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1894
1895              /* Update vectorial force */
1896             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1897             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1898             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1899
1900             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1901             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1902             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1903
1904             }
1905
1906             /**************************
1907              * CALCULATE INTERACTIONS *
1908              **************************/
1909
1910             if (gmx_mm_any_lt(rsq12,rcutoff2))
1911             {
1912
1913             /* REACTION-FIELD ELECTROSTATICS */
1914             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1915
1916             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1917
1918             fscal            = felec;
1919
1920             fscal            = _mm_and_ps(fscal,cutoff_mask);
1921
1922             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1923
1924              /* Update vectorial force */
1925             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1926             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1927             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1928
1929             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1930             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1931             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1932
1933             }
1934
1935             /**************************
1936              * CALCULATE INTERACTIONS *
1937              **************************/
1938
1939             if (gmx_mm_any_lt(rsq20,rcutoff2))
1940             {
1941
1942             /* REACTION-FIELD ELECTROSTATICS */
1943             felec            = _mm_mul_ps(qq20,_mm_msub_ps(rinv20,rinvsq20,krf2));
1944
1945             cutoff_mask      = _mm_cmplt_ps(rsq20,rcutoff2);
1946
1947             fscal            = felec;
1948
1949             fscal            = _mm_and_ps(fscal,cutoff_mask);
1950
1951             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1952
1953              /* Update vectorial force */
1954             fix2             = _mm_macc_ps(dx20,fscal,fix2);
1955             fiy2             = _mm_macc_ps(dy20,fscal,fiy2);
1956             fiz2             = _mm_macc_ps(dz20,fscal,fiz2);
1957
1958             fjx0             = _mm_macc_ps(dx20,fscal,fjx0);
1959             fjy0             = _mm_macc_ps(dy20,fscal,fjy0);
1960             fjz0             = _mm_macc_ps(dz20,fscal,fjz0);
1961
1962             }
1963
1964             /**************************
1965              * CALCULATE INTERACTIONS *
1966              **************************/
1967
1968             if (gmx_mm_any_lt(rsq21,rcutoff2))
1969             {
1970
1971             /* REACTION-FIELD ELECTROSTATICS */
1972             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1973
1974             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1975
1976             fscal            = felec;
1977
1978             fscal            = _mm_and_ps(fscal,cutoff_mask);
1979
1980             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1981
1982              /* Update vectorial force */
1983             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1984             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1985             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1986
1987             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1988             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1989             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1990
1991             }
1992
1993             /**************************
1994              * CALCULATE INTERACTIONS *
1995              **************************/
1996
1997             if (gmx_mm_any_lt(rsq22,rcutoff2))
1998             {
1999
2000             /* REACTION-FIELD ELECTROSTATICS */
2001             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
2002
2003             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
2004
2005             fscal            = felec;
2006
2007             fscal            = _mm_and_ps(fscal,cutoff_mask);
2008
2009             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2010
2011              /* Update vectorial force */
2012             fix2             = _mm_macc_ps(dx22,fscal,fix2);
2013             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
2014             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
2015
2016             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
2017             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
2018             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
2019
2020             }
2021
2022             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2023             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2024             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2025             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2026
2027             gmx_mm_decrement_3rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
2028                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
2029
2030             /* Inner loop uses 329 flops */
2031         }
2032
2033         /* End of innermost loop */
2034
2035         gmx_mm_update_iforce_3atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
2036                                               f+i_coord_offset,fshift+i_shift_offset);
2037
2038         /* Increment number of inner iterations */
2039         inneriter                  += j_index_end - j_index_start;
2040
2041         /* Outer loop uses 18 flops */
2042     }
2043
2044     /* Increment number of outer iterations */
2045     outeriter        += nri;
2046
2047     /* Update outer/inner flops */
2048
2049     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*329);
2050 }