src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecRFCut_VdwLJSh_GeomW4W4_avx_128_fma_single.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013, 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 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "types/simple.h"
  46 #include "vec.h"
  47 #include "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_VdwLJSh_GeomW4W4_VF_avx_128_fma_single
  54  * Electrostatics interaction: ReactionField
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Water4-Water4
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecRFCut_VdwLJSh_GeomW4W4_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              vdwioffset3;
  92     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  93     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  94     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  95     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  96     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  97     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  98     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  99     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
 100     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 101     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 102     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 103     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 104     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 105     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 106     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 107     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 108     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 109     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 110     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 111     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 112     real             *charge;
 113     int              nvdwtype;
 114     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 115     int              *vdwtype;
 116     real             *vdwparam;
 117     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 118     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 119     __m128           dummy_mask,cutoff_mask;
 120     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 121     __m128           one     = _mm_set1_ps(1.0);
 122     __m128           two     = _mm_set1_ps(2.0);
 123     x                = xx[0];
 124     f                = ff[0];
 125
 126     nri              = nlist->nri;
 127     iinr             = nlist->iinr;
 128     jindex           = nlist->jindex;
 129     jjnr             = nlist->jjnr;
 130     shiftidx         = nlist->shift;
 131     gid              = nlist->gid;
 132     shiftvec         = fr->shift_vec[0];
 133     fshift           = fr->fshift[0];
 134     facel            = _mm_set1_ps(fr->epsfac);
 135     charge           = mdatoms->chargeA;
 136     krf              = _mm_set1_ps(fr->ic->k_rf);
 137     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 138     crf              = _mm_set1_ps(fr->ic->c_rf);
 139     nvdwtype         = fr->ntype;
 140     vdwparam         = fr->nbfp;
 141     vdwtype          = mdatoms->typeA;
 142
 143     /* Setup water-specific parameters */
 144     inr              = nlist->iinr[0];
 145     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 146     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 147     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 148     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 149
 150     jq1              = _mm_set1_ps(charge[inr+1]);
 151     jq2              = _mm_set1_ps(charge[inr+2]);
 152     jq3              = _mm_set1_ps(charge[inr+3]);
 153     vdwjidx0A        = 2*vdwtype[inr+0];
 154     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 155     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 156     qq11             = _mm_mul_ps(iq1,jq1);
 157     qq12             = _mm_mul_ps(iq1,jq2);
 158     qq13             = _mm_mul_ps(iq1,jq3);
 159     qq21             = _mm_mul_ps(iq2,jq1);
 160     qq22             = _mm_mul_ps(iq2,jq2);
 161     qq23             = _mm_mul_ps(iq2,jq3);
 162     qq31             = _mm_mul_ps(iq3,jq1);
 163     qq32             = _mm_mul_ps(iq3,jq2);
 164     qq33             = _mm_mul_ps(iq3,jq3);
 165
 166     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 167     rcutoff_scalar   = fr->rcoulomb;
 168     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 169     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 170
 171     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 172     rvdw             = _mm_set1_ps(fr->rvdw);
 173
 174     /* Avoid stupid compiler warnings */
 175     jnrA = jnrB = jnrC = jnrD = 0;
 176     j_coord_offsetA = 0;
 177     j_coord_offsetB = 0;
 178     j_coord_offsetC = 0;
 179     j_coord_offsetD = 0;
 180
 181     outeriter        = 0;
 182     inneriter        = 0;
 183
 184     for(iidx=0;iidx<4*DIM;iidx++)
 185     {
 186         scratch[iidx] = 0.0;
 187     }
 188
 189     /* Start outer loop over neighborlists */
 190     for(iidx=0; iidx<nri; iidx++)
 191     {
 192         /* Load shift vector for this list */
 193         i_shift_offset   = DIM*shiftidx[iidx];
 194
 195         /* Load limits for loop over neighbors */
 196         j_index_start    = jindex[iidx];
 197         j_index_end      = jindex[iidx+1];
 198
 199         /* Get outer coordinate index */
 200         inr              = iinr[iidx];
 201         i_coord_offset   = DIM*inr;
 202
 203         /* Load i particle coords and add shift vector */
 204         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 205                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 206
 207         fix0             = _mm_setzero_ps();
 208         fiy0             = _mm_setzero_ps();
 209         fiz0             = _mm_setzero_ps();
 210         fix1             = _mm_setzero_ps();
 211         fiy1             = _mm_setzero_ps();
 212         fiz1             = _mm_setzero_ps();
 213         fix2             = _mm_setzero_ps();
 214         fiy2             = _mm_setzero_ps();
 215         fiz2             = _mm_setzero_ps();
 216         fix3             = _mm_setzero_ps();
 217         fiy3             = _mm_setzero_ps();
 218         fiz3             = _mm_setzero_ps();
 219
 220         /* Reset potential sums */
 221         velecsum         = _mm_setzero_ps();
 222         vvdwsum          = _mm_setzero_ps();
 223
 224         /* Start inner kernel loop */
 225         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 226         {
 227
 228             /* Get j neighbor index, and coordinate index */
 229             jnrA             = jjnr[jidx];
 230             jnrB             = jjnr[jidx+1];
 231             jnrC             = jjnr[jidx+2];
 232             jnrD             = jjnr[jidx+3];
 233             j_coord_offsetA  = DIM*jnrA;
 234             j_coord_offsetB  = DIM*jnrB;
 235             j_coord_offsetC  = DIM*jnrC;
 236             j_coord_offsetD  = DIM*jnrD;
 237
 238             /* load j atom coordinates */
 239             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 240                                               x+j_coord_offsetC,x+j_coord_offsetD,
 241                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 242                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 243
 244             /* Calculate displacement vector */
 245             dx00             = _mm_sub_ps(ix0,jx0);
 246             dy00             = _mm_sub_ps(iy0,jy0);
 247             dz00             = _mm_sub_ps(iz0,jz0);
 248             dx11             = _mm_sub_ps(ix1,jx1);
 249             dy11             = _mm_sub_ps(iy1,jy1);
 250             dz11             = _mm_sub_ps(iz1,jz1);
 251             dx12             = _mm_sub_ps(ix1,jx2);
 252             dy12             = _mm_sub_ps(iy1,jy2);
 253             dz12             = _mm_sub_ps(iz1,jz2);
 254             dx13             = _mm_sub_ps(ix1,jx3);
 255             dy13             = _mm_sub_ps(iy1,jy3);
 256             dz13             = _mm_sub_ps(iz1,jz3);
 257             dx21             = _mm_sub_ps(ix2,jx1);
 258             dy21             = _mm_sub_ps(iy2,jy1);
 259             dz21             = _mm_sub_ps(iz2,jz1);
 260             dx22             = _mm_sub_ps(ix2,jx2);
 261             dy22             = _mm_sub_ps(iy2,jy2);
 262             dz22             = _mm_sub_ps(iz2,jz2);
 263             dx23             = _mm_sub_ps(ix2,jx3);
 264             dy23             = _mm_sub_ps(iy2,jy3);
 265             dz23             = _mm_sub_ps(iz2,jz3);
 266             dx31             = _mm_sub_ps(ix3,jx1);
 267             dy31             = _mm_sub_ps(iy3,jy1);
 268             dz31             = _mm_sub_ps(iz3,jz1);
 269             dx32             = _mm_sub_ps(ix3,jx2);
 270             dy32             = _mm_sub_ps(iy3,jy2);
 271             dz32             = _mm_sub_ps(iz3,jz2);
 272             dx33             = _mm_sub_ps(ix3,jx3);
 273             dy33             = _mm_sub_ps(iy3,jy3);
 274             dz33             = _mm_sub_ps(iz3,jz3);
 275
 276             /* Calculate squared distance and things based on it */
 277             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 278             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 279             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 280             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 281             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 282             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 283             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 284             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 285             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 286             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 287
 288             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 289             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 290             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 291             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 292             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 293             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 294             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 295             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 296             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 297
 298             rinvsq00         = gmx_mm_inv_ps(rsq00);
 299             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 300             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 301             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 302             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 303             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 304             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 305             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 306             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 307             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 308
 309             fjx0             = _mm_setzero_ps();
 310             fjy0             = _mm_setzero_ps();
 311             fjz0             = _mm_setzero_ps();
 312             fjx1             = _mm_setzero_ps();
 313             fjy1             = _mm_setzero_ps();
 314             fjz1             = _mm_setzero_ps();
 315             fjx2             = _mm_setzero_ps();
 316             fjy2             = _mm_setzero_ps();
 317             fjz2             = _mm_setzero_ps();
 318             fjx3             = _mm_setzero_ps();
 319             fjy3             = _mm_setzero_ps();
 320             fjz3             = _mm_setzero_ps();
 321
 322             /**************************
 323              * CALCULATE INTERACTIONS *
 324              **************************/
 325
 326             if (gmx_mm_any_lt(rsq00,rcutoff2))
 327             {
 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(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 335                                           _mm_mul_ps( _mm_nmacc_ps(c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 336             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 337
 338             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 339
 340             /* Update potential sum for this i atom from the interaction with this j atom. */
 341             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 342             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 343
 344             fscal            = fvdw;
 345
 346             fscal            = _mm_and_ps(fscal,cutoff_mask);
 347
 348              /* Update vectorial force */
 349             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 350             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 351             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 352
 353             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 354             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 355             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 356
 357             }
 358
 359             /**************************
 360              * CALCULATE INTERACTIONS *
 361              **************************/
 362
 363             if (gmx_mm_any_lt(rsq11,rcutoff2))
 364             {
 365
 366             /* REACTION-FIELD ELECTROSTATICS */
 367             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 368             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 369
 370             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 371
 372             /* Update potential sum for this i atom from the interaction with this j atom. */
 373             velec            = _mm_and_ps(velec,cutoff_mask);
 374             velecsum         = _mm_add_ps(velecsum,velec);
 375
 376             fscal            = felec;
 377
 378             fscal            = _mm_and_ps(fscal,cutoff_mask);
 379
 380              /* Update vectorial force */
 381             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 382             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 383             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 384
 385             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 386             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 387             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 388
 389             }
 390
 391             /**************************
 392              * CALCULATE INTERACTIONS *
 393              **************************/
 394
 395             if (gmx_mm_any_lt(rsq12,rcutoff2))
 396             {
 397
 398             /* REACTION-FIELD ELECTROSTATICS */
 399             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 400             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 401
 402             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 403
 404             /* Update potential sum for this i atom from the interaction with this j atom. */
 405             velec            = _mm_and_ps(velec,cutoff_mask);
 406             velecsum         = _mm_add_ps(velecsum,velec);
 407
 408             fscal            = felec;
 409
 410             fscal            = _mm_and_ps(fscal,cutoff_mask);
 411
 412              /* Update vectorial force */
 413             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 414             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 415             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 416
 417             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 418             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 419             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 420
 421             }
 422
 423             /**************************
 424              * CALCULATE INTERACTIONS *
 425              **************************/
 426
 427             if (gmx_mm_any_lt(rsq13,rcutoff2))
 428             {
 429
 430             /* REACTION-FIELD ELECTROSTATICS */
 431             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_macc_ps(krf,rsq13,rinv13),crf));
 432             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
 433
 434             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
 435
 436             /* Update potential sum for this i atom from the interaction with this j atom. */
 437             velec            = _mm_and_ps(velec,cutoff_mask);
 438             velecsum         = _mm_add_ps(velecsum,velec);
 439
 440             fscal            = felec;
 441
 442             fscal            = _mm_and_ps(fscal,cutoff_mask);
 443
 444              /* Update vectorial force */
 445             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 446             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 447             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 448
 449             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 450             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 451             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 452
 453             }
 454
 455             /**************************
 456              * CALCULATE INTERACTIONS *
 457              **************************/
 458
 459             if (gmx_mm_any_lt(rsq21,rcutoff2))
 460             {
 461
 462             /* REACTION-FIELD ELECTROSTATICS */
 463             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 464             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 465
 466             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 467
 468             /* Update potential sum for this i atom from the interaction with this j atom. */
 469             velec            = _mm_and_ps(velec,cutoff_mask);
 470             velecsum         = _mm_add_ps(velecsum,velec);
 471
 472             fscal            = felec;
 473
 474             fscal            = _mm_and_ps(fscal,cutoff_mask);
 475
 476              /* Update vectorial force */
 477             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 478             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 479             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 480
 481             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 482             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 483             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 484
 485             }
 486
 487             /**************************
 488              * CALCULATE INTERACTIONS *
 489              **************************/
 490
 491             if (gmx_mm_any_lt(rsq22,rcutoff2))
 492             {
 493
 494             /* REACTION-FIELD ELECTROSTATICS */
 495             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 496             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 497
 498             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 499
 500             /* Update potential sum for this i atom from the interaction with this j atom. */
 501             velec            = _mm_and_ps(velec,cutoff_mask);
 502             velecsum         = _mm_add_ps(velecsum,velec);
 503
 504             fscal            = felec;
 505
 506             fscal            = _mm_and_ps(fscal,cutoff_mask);
 507
 508              /* Update vectorial force */
 509             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 510             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 511             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 512
 513             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 514             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 515             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 516
 517             }
 518
 519             /**************************
 520              * CALCULATE INTERACTIONS *
 521              **************************/
 522
 523             if (gmx_mm_any_lt(rsq23,rcutoff2))
 524             {
 525
 526             /* REACTION-FIELD ELECTROSTATICS */
 527             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_macc_ps(krf,rsq23,rinv23),crf));
 528             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
 529
 530             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
 531
 532             /* Update potential sum for this i atom from the interaction with this j atom. */
 533             velec            = _mm_and_ps(velec,cutoff_mask);
 534             velecsum         = _mm_add_ps(velecsum,velec);
 535
 536             fscal            = felec;
 537
 538             fscal            = _mm_and_ps(fscal,cutoff_mask);
 539
 540              /* Update vectorial force */
 541             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 542             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 543             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 544
 545             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 546             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 547             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 548
 549             }
 550
 551             /**************************
 552              * CALCULATE INTERACTIONS *
 553              **************************/
 554
 555             if (gmx_mm_any_lt(rsq31,rcutoff2))
 556             {
 557
 558             /* REACTION-FIELD ELECTROSTATICS */
 559             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_macc_ps(krf,rsq31,rinv31),crf));
 560             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
 561
 562             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
 563
 564             /* Update potential sum for this i atom from the interaction with this j atom. */
 565             velec            = _mm_and_ps(velec,cutoff_mask);
 566             velecsum         = _mm_add_ps(velecsum,velec);
 567
 568             fscal            = felec;
 569
 570             fscal            = _mm_and_ps(fscal,cutoff_mask);
 571
 572              /* Update vectorial force */
 573             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 574             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 575             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 576
 577             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 578             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 579             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 580
 581             }
 582
 583             /**************************
 584              * CALCULATE INTERACTIONS *
 585              **************************/
 586
 587             if (gmx_mm_any_lt(rsq32,rcutoff2))
 588             {
 589
 590             /* REACTION-FIELD ELECTROSTATICS */
 591             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_macc_ps(krf,rsq32,rinv32),crf));
 592             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
 593
 594             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
 595
 596             /* Update potential sum for this i atom from the interaction with this j atom. */
 597             velec            = _mm_and_ps(velec,cutoff_mask);
 598             velecsum         = _mm_add_ps(velecsum,velec);
 599
 600             fscal            = felec;
 601
 602             fscal            = _mm_and_ps(fscal,cutoff_mask);
 603
 604              /* Update vectorial force */
 605             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 606             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 607             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 608
 609             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 610             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 611             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 612
 613             }
 614
 615             /**************************
 616              * CALCULATE INTERACTIONS *
 617              **************************/
 618
 619             if (gmx_mm_any_lt(rsq33,rcutoff2))
 620             {
 621
 622             /* REACTION-FIELD ELECTROSTATICS */
 623             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_macc_ps(krf,rsq33,rinv33),crf));
 624             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
 625
 626             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
 627
 628             /* Update potential sum for this i atom from the interaction with this j atom. */
 629             velec            = _mm_and_ps(velec,cutoff_mask);
 630             velecsum         = _mm_add_ps(velecsum,velec);
 631
 632             fscal            = felec;
 633
 634             fscal            = _mm_and_ps(fscal,cutoff_mask);
 635
 636              /* Update vectorial force */
 637             fix3             = _mm_macc_ps(dx33,fscal,fix3);
 638             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
 639             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
 640
 641             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
 642             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
 643             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
 644
 645             }
 646
 647             fjptrA             = f+j_coord_offsetA;
 648             fjptrB             = f+j_coord_offsetB;
 649             fjptrC             = f+j_coord_offsetC;
 650             fjptrD             = f+j_coord_offsetD;
 651
 652             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 653                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 654                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 655
 656             /* Inner loop uses 398 flops */
 657         }
 658
 659         if(jidx<j_index_end)
 660         {
 661
 662             /* Get j neighbor index, and coordinate index */
 663             jnrlistA         = jjnr[jidx];
 664             jnrlistB         = jjnr[jidx+1];
 665             jnrlistC         = jjnr[jidx+2];
 666             jnrlistD         = jjnr[jidx+3];
 667             /* Sign of each element will be negative for non-real atoms.
 668              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 669              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 670              */
 671             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 672             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 673             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 674             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 675             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 676             j_coord_offsetA  = DIM*jnrA;
 677             j_coord_offsetB  = DIM*jnrB;
 678             j_coord_offsetC  = DIM*jnrC;
 679             j_coord_offsetD  = DIM*jnrD;
 680
 681             /* load j atom coordinates */
 682             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 683                                               x+j_coord_offsetC,x+j_coord_offsetD,
 684                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 685                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 686
 687             /* Calculate displacement vector */
 688             dx00             = _mm_sub_ps(ix0,jx0);
 689             dy00             = _mm_sub_ps(iy0,jy0);
 690             dz00             = _mm_sub_ps(iz0,jz0);
 691             dx11             = _mm_sub_ps(ix1,jx1);
 692             dy11             = _mm_sub_ps(iy1,jy1);
 693             dz11             = _mm_sub_ps(iz1,jz1);
 694             dx12             = _mm_sub_ps(ix1,jx2);
 695             dy12             = _mm_sub_ps(iy1,jy2);
 696             dz12             = _mm_sub_ps(iz1,jz2);
 697             dx13             = _mm_sub_ps(ix1,jx3);
 698             dy13             = _mm_sub_ps(iy1,jy3);
 699             dz13             = _mm_sub_ps(iz1,jz3);
 700             dx21             = _mm_sub_ps(ix2,jx1);
 701             dy21             = _mm_sub_ps(iy2,jy1);
 702             dz21             = _mm_sub_ps(iz2,jz1);
 703             dx22             = _mm_sub_ps(ix2,jx2);
 704             dy22             = _mm_sub_ps(iy2,jy2);
 705             dz22             = _mm_sub_ps(iz2,jz2);
 706             dx23             = _mm_sub_ps(ix2,jx3);
 707             dy23             = _mm_sub_ps(iy2,jy3);
 708             dz23             = _mm_sub_ps(iz2,jz3);
 709             dx31             = _mm_sub_ps(ix3,jx1);
 710             dy31             = _mm_sub_ps(iy3,jy1);
 711             dz31             = _mm_sub_ps(iz3,jz1);
 712             dx32             = _mm_sub_ps(ix3,jx2);
 713             dy32             = _mm_sub_ps(iy3,jy2);
 714             dz32             = _mm_sub_ps(iz3,jz2);
 715             dx33             = _mm_sub_ps(ix3,jx3);
 716             dy33             = _mm_sub_ps(iy3,jy3);
 717             dz33             = _mm_sub_ps(iz3,jz3);
 718
 719             /* Calculate squared distance and things based on it */
 720             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 721             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 722             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 723             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 724             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 725             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 726             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 727             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 728             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 729             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 730
 731             rinv11           = gmx_mm_invsqrt_ps(rsq11);
 732             rinv12           = gmx_mm_invsqrt_ps(rsq12);
 733             rinv13           = gmx_mm_invsqrt_ps(rsq13);
 734             rinv21           = gmx_mm_invsqrt_ps(rsq21);
 735             rinv22           = gmx_mm_invsqrt_ps(rsq22);
 736             rinv23           = gmx_mm_invsqrt_ps(rsq23);
 737             rinv31           = gmx_mm_invsqrt_ps(rsq31);
 738             rinv32           = gmx_mm_invsqrt_ps(rsq32);
 739             rinv33           = gmx_mm_invsqrt_ps(rsq33);
 740
 741             rinvsq00         = gmx_mm_inv_ps(rsq00);
 742             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 743             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 744             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 745             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 746             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 747             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 748             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 749             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 750             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 751
 752             fjx0             = _mm_setzero_ps();
 753             fjy0             = _mm_setzero_ps();
 754             fjz0             = _mm_setzero_ps();
 755             fjx1             = _mm_setzero_ps();
 756             fjy1             = _mm_setzero_ps();
 757             fjz1             = _mm_setzero_ps();
 758             fjx2             = _mm_setzero_ps();
 759             fjy2             = _mm_setzero_ps();
 760             fjz2             = _mm_setzero_ps();
 761             fjx3             = _mm_setzero_ps();
 762             fjy3             = _mm_setzero_ps();
 763             fjz3             = _mm_setzero_ps();
 764
 765             /**************************
 766              * CALCULATE INTERACTIONS *
 767              **************************/
 768
 769             if (gmx_mm_any_lt(rsq00,rcutoff2))
 770             {
 771
 772             /* LENNARD-JONES DISPERSION/REPULSION */
 773
 774             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 775             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 776             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 777             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 778                                           _mm_mul_ps( _mm_nmacc_ps(c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 779             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 780
 781             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 782
 783             /* Update potential sum for this i atom from the interaction with this j atom. */
 784             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 785             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 786             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 787
 788             fscal            = fvdw;
 789
 790             fscal            = _mm_and_ps(fscal,cutoff_mask);
 791
 792             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 793
 794              /* Update vectorial force */
 795             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 796             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 797             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 798
 799             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 800             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 801             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 802
 803             }
 804
 805             /**************************
 806              * CALCULATE INTERACTIONS *
 807              **************************/
 808
 809             if (gmx_mm_any_lt(rsq11,rcutoff2))
 810             {
 811
 812             /* REACTION-FIELD ELECTROSTATICS */
 813             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 814             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 815
 816             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 817
 818             /* Update potential sum for this i atom from the interaction with this j atom. */
 819             velec            = _mm_and_ps(velec,cutoff_mask);
 820             velec            = _mm_andnot_ps(dummy_mask,velec);
 821             velecsum         = _mm_add_ps(velecsum,velec);
 822
 823             fscal            = felec;
 824
 825             fscal            = _mm_and_ps(fscal,cutoff_mask);
 826
 827             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 828
 829              /* Update vectorial force */
 830             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 831             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 832             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 833
 834             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 835             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 836             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 837
 838             }
 839
 840             /**************************
 841              * CALCULATE INTERACTIONS *
 842              **************************/
 843
 844             if (gmx_mm_any_lt(rsq12,rcutoff2))
 845             {
 846
 847             /* REACTION-FIELD ELECTROSTATICS */
 848             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 849             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 850
 851             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 852
 853             /* Update potential sum for this i atom from the interaction with this j atom. */
 854             velec            = _mm_and_ps(velec,cutoff_mask);
 855             velec            = _mm_andnot_ps(dummy_mask,velec);
 856             velecsum         = _mm_add_ps(velecsum,velec);
 857
 858             fscal            = felec;
 859
 860             fscal            = _mm_and_ps(fscal,cutoff_mask);
 861
 862             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 863
 864              /* Update vectorial force */
 865             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 866             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 867             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 868
 869             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 870             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 871             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 872
 873             }
 874
 875             /**************************
 876              * CALCULATE INTERACTIONS *
 877              **************************/
 878
 879             if (gmx_mm_any_lt(rsq13,rcutoff2))
 880             {
 881
 882             /* REACTION-FIELD ELECTROSTATICS */
 883             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_macc_ps(krf,rsq13,rinv13),crf));
 884             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
 885
 886             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
 887
 888             /* Update potential sum for this i atom from the interaction with this j atom. */
 889             velec            = _mm_and_ps(velec,cutoff_mask);
 890             velec            = _mm_andnot_ps(dummy_mask,velec);
 891             velecsum         = _mm_add_ps(velecsum,velec);
 892
 893             fscal            = felec;
 894
 895             fscal            = _mm_and_ps(fscal,cutoff_mask);
 896
 897             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 898
 899              /* Update vectorial force */
 900             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 901             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 902             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 903
 904             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 905             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 906             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 907
 908             }
 909
 910             /**************************
 911              * CALCULATE INTERACTIONS *
 912              **************************/
 913
 914             if (gmx_mm_any_lt(rsq21,rcutoff2))
 915             {
 916
 917             /* REACTION-FIELD ELECTROSTATICS */
 918             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 919             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 920
 921             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 922
 923             /* Update potential sum for this i atom from the interaction with this j atom. */
 924             velec            = _mm_and_ps(velec,cutoff_mask);
 925             velec            = _mm_andnot_ps(dummy_mask,velec);
 926             velecsum         = _mm_add_ps(velecsum,velec);
 927
 928             fscal            = felec;
 929
 930             fscal            = _mm_and_ps(fscal,cutoff_mask);
 931
 932             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 933
 934              /* Update vectorial force */
 935             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 936             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 937             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 938
 939             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 940             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 941             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 942
 943             }
 944
 945             /**************************
 946              * CALCULATE INTERACTIONS *
 947              **************************/
 948
 949             if (gmx_mm_any_lt(rsq22,rcutoff2))
 950             {
 951
 952             /* REACTION-FIELD ELECTROSTATICS */
 953             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 954             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 955
 956             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 957
 958             /* Update potential sum for this i atom from the interaction with this j atom. */
 959             velec            = _mm_and_ps(velec,cutoff_mask);
 960             velec            = _mm_andnot_ps(dummy_mask,velec);
 961             velecsum         = _mm_add_ps(velecsum,velec);
 962
 963             fscal            = felec;
 964
 965             fscal            = _mm_and_ps(fscal,cutoff_mask);
 966
 967             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 968
 969              /* Update vectorial force */
 970             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 971             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 972             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 973
 974             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 975             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 976             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 977
 978             }
 979
 980             /**************************
 981              * CALCULATE INTERACTIONS *
 982              **************************/
 983
 984             if (gmx_mm_any_lt(rsq23,rcutoff2))
 985             {
 986
 987             /* REACTION-FIELD ELECTROSTATICS */
 988             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_macc_ps(krf,rsq23,rinv23),crf));
 989             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
 990
 991             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
 992
 993             /* Update potential sum for this i atom from the interaction with this j atom. */
 994             velec            = _mm_and_ps(velec,cutoff_mask);
 995             velec            = _mm_andnot_ps(dummy_mask,velec);
 996             velecsum         = _mm_add_ps(velecsum,velec);
 997
 998             fscal            = felec;
 999
1000             fscal            = _mm_and_ps(fscal,cutoff_mask);
1001
1002             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1003
1004              /* Update vectorial force */
1005             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1006             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1007             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1008
1009             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1010             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1011             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1012
1013             }
1014
1015             /**************************
1016              * CALCULATE INTERACTIONS *
1017              **************************/
1018
1019             if (gmx_mm_any_lt(rsq31,rcutoff2))
1020             {
1021
1022             /* REACTION-FIELD ELECTROSTATICS */
1023             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_macc_ps(krf,rsq31,rinv31),crf));
1024             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
1025
1026             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
1027
1028             /* Update potential sum for this i atom from the interaction with this j atom. */
1029             velec            = _mm_and_ps(velec,cutoff_mask);
1030             velec            = _mm_andnot_ps(dummy_mask,velec);
1031             velecsum         = _mm_add_ps(velecsum,velec);
1032
1033             fscal            = felec;
1034
1035             fscal            = _mm_and_ps(fscal,cutoff_mask);
1036
1037             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1038
1039              /* Update vectorial force */
1040             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1041             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1042             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1043
1044             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1045             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1046             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1047
1048             }
1049
1050             /**************************
1051              * CALCULATE INTERACTIONS *
1052              **************************/
1053
1054             if (gmx_mm_any_lt(rsq32,rcutoff2))
1055             {
1056
1057             /* REACTION-FIELD ELECTROSTATICS */
1058             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_macc_ps(krf,rsq32,rinv32),crf));
1059             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
1060
1061             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1062
1063             /* Update potential sum for this i atom from the interaction with this j atom. */
1064             velec            = _mm_and_ps(velec,cutoff_mask);
1065             velec            = _mm_andnot_ps(dummy_mask,velec);
1066             velecsum         = _mm_add_ps(velecsum,velec);
1067
1068             fscal            = felec;
1069
1070             fscal            = _mm_and_ps(fscal,cutoff_mask);
1071
1072             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1073
1074              /* Update vectorial force */
1075             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1076             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1077             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1078
1079             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1080             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1081             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1082
1083             }
1084
1085             /**************************
1086              * CALCULATE INTERACTIONS *
1087              **************************/
1088
1089             if (gmx_mm_any_lt(rsq33,rcutoff2))
1090             {
1091
1092             /* REACTION-FIELD ELECTROSTATICS */
1093             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_macc_ps(krf,rsq33,rinv33),crf));
1094             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
1095
1096             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
1097
1098             /* Update potential sum for this i atom from the interaction with this j atom. */
1099             velec            = _mm_and_ps(velec,cutoff_mask);
1100             velec            = _mm_andnot_ps(dummy_mask,velec);
1101             velecsum         = _mm_add_ps(velecsum,velec);
1102
1103             fscal            = felec;
1104
1105             fscal            = _mm_and_ps(fscal,cutoff_mask);
1106
1107             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1108
1109              /* Update vectorial force */
1110             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1111             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1112             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1113
1114             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1115             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1116             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1117
1118             }
1119
1120             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1121             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1122             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1123             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1124
1125             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1126                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1127                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1128
1129             /* Inner loop uses 398 flops */
1130         }
1131
1132         /* End of innermost loop */
1133
1134         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1135                                               f+i_coord_offset,fshift+i_shift_offset);
1136
1137         ggid                        = gid[iidx];
1138         /* Update potential energies */
1139         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1140         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1141
1142         /* Increment number of inner iterations */
1143         inneriter                  += j_index_end - j_index_start;
1144
1145         /* Outer loop uses 26 flops */
1146     }
1147
1148     /* Increment number of outer iterations */
1149     outeriter        += nri;
1150
1151     /* Update outer/inner flops */
1152
1153     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*398);
1154 }
1155 /*
1156  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW4W4_F_avx_128_fma_single
1157  * Electrostatics interaction: ReactionField
1158  * VdW interaction:            LennardJones
1159  * Geometry:                   Water4-Water4
1160  * Calculate force/pot:        Force
1161  */
1162 void
1163 nb_kernel_ElecRFCut_VdwLJSh_GeomW4W4_F_avx_128_fma_single
1164                     (t_nblist                    * gmx_restrict       nlist,
1165                      rvec                        * gmx_restrict          xx,
1166                      rvec                        * gmx_restrict          ff,
1167                      t_forcerec                  * gmx_restrict          fr,
1168                      t_mdatoms                   * gmx_restrict     mdatoms,
1169                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1170                      t_nrnb                      * gmx_restrict        nrnb)
1171 {
1172     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1173      * just 0 for non-waters.
1174      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1175      * jnr indices corresponding to data put in the four positions in the SIMD register.
1176      */
1177     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1178     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1179     int              jnrA,jnrB,jnrC,jnrD;
1180     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1181     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1182     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1183     real             rcutoff_scalar;
1184     real             *shiftvec,*fshift,*x,*f;
1185     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1186     real             scratch[4*DIM];
1187     __m128           fscal,rcutoff,rcutoff2,jidxall;
1188     int              vdwioffset0;
1189     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1190     int              vdwioffset1;
1191     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1192     int              vdwioffset2;
1193     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1194     int              vdwioffset3;
1195     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1196     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1197     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1198     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1199     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1200     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1201     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1202     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1203     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1204     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1205     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1206     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1207     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1208     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1209     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1210     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1211     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1212     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1213     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1214     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1215     real             *charge;
1216     int              nvdwtype;
1217     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1218     int              *vdwtype;
1219     real             *vdwparam;
1220     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1221     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1222     __m128           dummy_mask,cutoff_mask;
1223     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1224     __m128           one     = _mm_set1_ps(1.0);
1225     __m128           two     = _mm_set1_ps(2.0);
1226     x                = xx[0];
1227     f                = ff[0];
1228
1229     nri              = nlist->nri;
1230     iinr             = nlist->iinr;
1231     jindex           = nlist->jindex;
1232     jjnr             = nlist->jjnr;
1233     shiftidx         = nlist->shift;
1234     gid              = nlist->gid;
1235     shiftvec         = fr->shift_vec[0];
1236     fshift           = fr->fshift[0];
1237     facel            = _mm_set1_ps(fr->epsfac);
1238     charge           = mdatoms->chargeA;
1239     krf              = _mm_set1_ps(fr->ic->k_rf);
1240     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1241     crf              = _mm_set1_ps(fr->ic->c_rf);
1242     nvdwtype         = fr->ntype;
1243     vdwparam         = fr->nbfp;
1244     vdwtype          = mdatoms->typeA;
1245
1246     /* Setup water-specific parameters */
1247     inr              = nlist->iinr[0];
1248     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1249     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1250     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1251     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1252
1253     jq1              = _mm_set1_ps(charge[inr+1]);
1254     jq2              = _mm_set1_ps(charge[inr+2]);
1255     jq3              = _mm_set1_ps(charge[inr+3]);
1256     vdwjidx0A        = 2*vdwtype[inr+0];
1257     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1258     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1259     qq11             = _mm_mul_ps(iq1,jq1);
1260     qq12             = _mm_mul_ps(iq1,jq2);
1261     qq13             = _mm_mul_ps(iq1,jq3);
1262     qq21             = _mm_mul_ps(iq2,jq1);
1263     qq22             = _mm_mul_ps(iq2,jq2);
1264     qq23             = _mm_mul_ps(iq2,jq3);
1265     qq31             = _mm_mul_ps(iq3,jq1);
1266     qq32             = _mm_mul_ps(iq3,jq2);
1267     qq33             = _mm_mul_ps(iq3,jq3);
1268
1269     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1270     rcutoff_scalar   = fr->rcoulomb;
1271     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1272     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1273
1274     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
1275     rvdw             = _mm_set1_ps(fr->rvdw);
1276
1277     /* Avoid stupid compiler warnings */
1278     jnrA = jnrB = jnrC = jnrD = 0;
1279     j_coord_offsetA = 0;
1280     j_coord_offsetB = 0;
1281     j_coord_offsetC = 0;
1282     j_coord_offsetD = 0;
1283
1284     outeriter        = 0;
1285     inneriter        = 0;
1286
1287     for(iidx=0;iidx<4*DIM;iidx++)
1288     {
1289         scratch[iidx] = 0.0;
1290     }
1291
1292     /* Start outer loop over neighborlists */
1293     for(iidx=0; iidx<nri; iidx++)
1294     {
1295         /* Load shift vector for this list */
1296         i_shift_offset   = DIM*shiftidx[iidx];
1297
1298         /* Load limits for loop over neighbors */
1299         j_index_start    = jindex[iidx];
1300         j_index_end      = jindex[iidx+1];
1301
1302         /* Get outer coordinate index */
1303         inr              = iinr[iidx];
1304         i_coord_offset   = DIM*inr;
1305
1306         /* Load i particle coords and add shift vector */
1307         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1308                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1309
1310         fix0             = _mm_setzero_ps();
1311         fiy0             = _mm_setzero_ps();
1312         fiz0             = _mm_setzero_ps();
1313         fix1             = _mm_setzero_ps();
1314         fiy1             = _mm_setzero_ps();
1315         fiz1             = _mm_setzero_ps();
1316         fix2             = _mm_setzero_ps();
1317         fiy2             = _mm_setzero_ps();
1318         fiz2             = _mm_setzero_ps();
1319         fix3             = _mm_setzero_ps();
1320         fiy3             = _mm_setzero_ps();
1321         fiz3             = _mm_setzero_ps();
1322
1323         /* Start inner kernel loop */
1324         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1325         {
1326
1327             /* Get j neighbor index, and coordinate index */
1328             jnrA             = jjnr[jidx];
1329             jnrB             = jjnr[jidx+1];
1330             jnrC             = jjnr[jidx+2];
1331             jnrD             = jjnr[jidx+3];
1332             j_coord_offsetA  = DIM*jnrA;
1333             j_coord_offsetB  = DIM*jnrB;
1334             j_coord_offsetC  = DIM*jnrC;
1335             j_coord_offsetD  = DIM*jnrD;
1336
1337             /* load j atom coordinates */
1338             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1339                                               x+j_coord_offsetC,x+j_coord_offsetD,
1340                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1341                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1342
1343             /* Calculate displacement vector */
1344             dx00             = _mm_sub_ps(ix0,jx0);
1345             dy00             = _mm_sub_ps(iy0,jy0);
1346             dz00             = _mm_sub_ps(iz0,jz0);
1347             dx11             = _mm_sub_ps(ix1,jx1);
1348             dy11             = _mm_sub_ps(iy1,jy1);
1349             dz11             = _mm_sub_ps(iz1,jz1);
1350             dx12             = _mm_sub_ps(ix1,jx2);
1351             dy12             = _mm_sub_ps(iy1,jy2);
1352             dz12             = _mm_sub_ps(iz1,jz2);
1353             dx13             = _mm_sub_ps(ix1,jx3);
1354             dy13             = _mm_sub_ps(iy1,jy3);
1355             dz13             = _mm_sub_ps(iz1,jz3);
1356             dx21             = _mm_sub_ps(ix2,jx1);
1357             dy21             = _mm_sub_ps(iy2,jy1);
1358             dz21             = _mm_sub_ps(iz2,jz1);
1359             dx22             = _mm_sub_ps(ix2,jx2);
1360             dy22             = _mm_sub_ps(iy2,jy2);
1361             dz22             = _mm_sub_ps(iz2,jz2);
1362             dx23             = _mm_sub_ps(ix2,jx3);
1363             dy23             = _mm_sub_ps(iy2,jy3);
1364             dz23             = _mm_sub_ps(iz2,jz3);
1365             dx31             = _mm_sub_ps(ix3,jx1);
1366             dy31             = _mm_sub_ps(iy3,jy1);
1367             dz31             = _mm_sub_ps(iz3,jz1);
1368             dx32             = _mm_sub_ps(ix3,jx2);
1369             dy32             = _mm_sub_ps(iy3,jy2);
1370             dz32             = _mm_sub_ps(iz3,jz2);
1371             dx33             = _mm_sub_ps(ix3,jx3);
1372             dy33             = _mm_sub_ps(iy3,jy3);
1373             dz33             = _mm_sub_ps(iz3,jz3);
1374
1375             /* Calculate squared distance and things based on it */
1376             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1377             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1378             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1379             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1380             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1381             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1382             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1383             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1384             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1385             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1386
1387             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1388             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1389             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1390             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1391             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1392             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1393             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1394             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1395             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1396
1397             rinvsq00         = gmx_mm_inv_ps(rsq00);
1398             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1399             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1400             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1401             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1402             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1403             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1404             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1405             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1406             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1407
1408             fjx0             = _mm_setzero_ps();
1409             fjy0             = _mm_setzero_ps();
1410             fjz0             = _mm_setzero_ps();
1411             fjx1             = _mm_setzero_ps();
1412             fjy1             = _mm_setzero_ps();
1413             fjz1             = _mm_setzero_ps();
1414             fjx2             = _mm_setzero_ps();
1415             fjy2             = _mm_setzero_ps();
1416             fjz2             = _mm_setzero_ps();
1417             fjx3             = _mm_setzero_ps();
1418             fjy3             = _mm_setzero_ps();
1419             fjz3             = _mm_setzero_ps();
1420
1421             /**************************
1422              * CALCULATE INTERACTIONS *
1423              **************************/
1424
1425             if (gmx_mm_any_lt(rsq00,rcutoff2))
1426             {
1427
1428             /* LENNARD-JONES DISPERSION/REPULSION */
1429
1430             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1431             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1432
1433             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1434
1435             fscal            = fvdw;
1436
1437             fscal            = _mm_and_ps(fscal,cutoff_mask);
1438
1439              /* Update vectorial force */
1440             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1441             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1442             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1443
1444             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1445             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1446             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1447
1448             }
1449
1450             /**************************
1451              * CALCULATE INTERACTIONS *
1452              **************************/
1453
1454             if (gmx_mm_any_lt(rsq11,rcutoff2))
1455             {
1456
1457             /* REACTION-FIELD ELECTROSTATICS */
1458             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1459
1460             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1461
1462             fscal            = felec;
1463
1464             fscal            = _mm_and_ps(fscal,cutoff_mask);
1465
1466              /* Update vectorial force */
1467             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1468             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1469             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1470
1471             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1472             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1473             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1474
1475             }
1476
1477             /**************************
1478              * CALCULATE INTERACTIONS *
1479              **************************/
1480
1481             if (gmx_mm_any_lt(rsq12,rcutoff2))
1482             {
1483
1484             /* REACTION-FIELD ELECTROSTATICS */
1485             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1486
1487             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1488
1489             fscal            = felec;
1490
1491             fscal            = _mm_and_ps(fscal,cutoff_mask);
1492
1493              /* Update vectorial force */
1494             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1495             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1496             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1497
1498             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1499             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1500             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1501
1502             }
1503
1504             /**************************
1505              * CALCULATE INTERACTIONS *
1506              **************************/
1507
1508             if (gmx_mm_any_lt(rsq13,rcutoff2))
1509             {
1510
1511             /* REACTION-FIELD ELECTROSTATICS */
1512             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
1513
1514             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
1515
1516             fscal            = felec;
1517
1518             fscal            = _mm_and_ps(fscal,cutoff_mask);
1519
1520              /* Update vectorial force */
1521             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1522             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1523             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1524
1525             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1526             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1527             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1528
1529             }
1530
1531             /**************************
1532              * CALCULATE INTERACTIONS *
1533              **************************/
1534
1535             if (gmx_mm_any_lt(rsq21,rcutoff2))
1536             {
1537
1538             /* REACTION-FIELD ELECTROSTATICS */
1539             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1540
1541             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1542
1543             fscal            = felec;
1544
1545             fscal            = _mm_and_ps(fscal,cutoff_mask);
1546
1547              /* Update vectorial force */
1548             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1549             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1550             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1551
1552             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1553             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1554             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1555
1556             }
1557
1558             /**************************
1559              * CALCULATE INTERACTIONS *
1560              **************************/
1561
1562             if (gmx_mm_any_lt(rsq22,rcutoff2))
1563             {
1564
1565             /* REACTION-FIELD ELECTROSTATICS */
1566             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1567
1568             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1569
1570             fscal            = felec;
1571
1572             fscal            = _mm_and_ps(fscal,cutoff_mask);
1573
1574              /* Update vectorial force */
1575             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1576             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1577             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1578
1579             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1580             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1581             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1582
1583             }
1584
1585             /**************************
1586              * CALCULATE INTERACTIONS *
1587              **************************/
1588
1589             if (gmx_mm_any_lt(rsq23,rcutoff2))
1590             {
1591
1592             /* REACTION-FIELD ELECTROSTATICS */
1593             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
1594
1595             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
1596
1597             fscal            = felec;
1598
1599             fscal            = _mm_and_ps(fscal,cutoff_mask);
1600
1601              /* Update vectorial force */
1602             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1603             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1604             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1605
1606             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1607             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1608             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1609
1610             }
1611
1612             /**************************
1613              * CALCULATE INTERACTIONS *
1614              **************************/
1615
1616             if (gmx_mm_any_lt(rsq31,rcutoff2))
1617             {
1618
1619             /* REACTION-FIELD ELECTROSTATICS */
1620             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
1621
1622             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
1623
1624             fscal            = felec;
1625
1626             fscal            = _mm_and_ps(fscal,cutoff_mask);
1627
1628              /* Update vectorial force */
1629             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1630             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1631             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1632
1633             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1634             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1635             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1636
1637             }
1638
1639             /**************************
1640              * CALCULATE INTERACTIONS *
1641              **************************/
1642
1643             if (gmx_mm_any_lt(rsq32,rcutoff2))
1644             {
1645
1646             /* REACTION-FIELD ELECTROSTATICS */
1647             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
1648
1649             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1650
1651             fscal            = felec;
1652
1653             fscal            = _mm_and_ps(fscal,cutoff_mask);
1654
1655              /* Update vectorial force */
1656             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1657             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1658             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1659
1660             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1661             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1662             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1663
1664             }
1665
1666             /**************************
1667              * CALCULATE INTERACTIONS *
1668              **************************/
1669
1670             if (gmx_mm_any_lt(rsq33,rcutoff2))
1671             {
1672
1673             /* REACTION-FIELD ELECTROSTATICS */
1674             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
1675
1676             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
1677
1678             fscal            = felec;
1679
1680             fscal            = _mm_and_ps(fscal,cutoff_mask);
1681
1682              /* Update vectorial force */
1683             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1684             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1685             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1686
1687             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1688             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1689             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1690
1691             }
1692
1693             fjptrA             = f+j_coord_offsetA;
1694             fjptrB             = f+j_coord_offsetB;
1695             fjptrC             = f+j_coord_offsetC;
1696             fjptrD             = f+j_coord_offsetD;
1697
1698             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1699                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1700                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1701
1702             /* Inner loop uses 333 flops */
1703         }
1704
1705         if(jidx<j_index_end)
1706         {
1707
1708             /* Get j neighbor index, and coordinate index */
1709             jnrlistA         = jjnr[jidx];
1710             jnrlistB         = jjnr[jidx+1];
1711             jnrlistC         = jjnr[jidx+2];
1712             jnrlistD         = jjnr[jidx+3];
1713             /* Sign of each element will be negative for non-real atoms.
1714              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1715              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1716              */
1717             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1718             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1719             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1720             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1721             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1722             j_coord_offsetA  = DIM*jnrA;
1723             j_coord_offsetB  = DIM*jnrB;
1724             j_coord_offsetC  = DIM*jnrC;
1725             j_coord_offsetD  = DIM*jnrD;
1726
1727             /* load j atom coordinates */
1728             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1729                                               x+j_coord_offsetC,x+j_coord_offsetD,
1730                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1731                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1732
1733             /* Calculate displacement vector */
1734             dx00             = _mm_sub_ps(ix0,jx0);
1735             dy00             = _mm_sub_ps(iy0,jy0);
1736             dz00             = _mm_sub_ps(iz0,jz0);
1737             dx11             = _mm_sub_ps(ix1,jx1);
1738             dy11             = _mm_sub_ps(iy1,jy1);
1739             dz11             = _mm_sub_ps(iz1,jz1);
1740             dx12             = _mm_sub_ps(ix1,jx2);
1741             dy12             = _mm_sub_ps(iy1,jy2);
1742             dz12             = _mm_sub_ps(iz1,jz2);
1743             dx13             = _mm_sub_ps(ix1,jx3);
1744             dy13             = _mm_sub_ps(iy1,jy3);
1745             dz13             = _mm_sub_ps(iz1,jz3);
1746             dx21             = _mm_sub_ps(ix2,jx1);
1747             dy21             = _mm_sub_ps(iy2,jy1);
1748             dz21             = _mm_sub_ps(iz2,jz1);
1749             dx22             = _mm_sub_ps(ix2,jx2);
1750             dy22             = _mm_sub_ps(iy2,jy2);
1751             dz22             = _mm_sub_ps(iz2,jz2);
1752             dx23             = _mm_sub_ps(ix2,jx3);
1753             dy23             = _mm_sub_ps(iy2,jy3);
1754             dz23             = _mm_sub_ps(iz2,jz3);
1755             dx31             = _mm_sub_ps(ix3,jx1);
1756             dy31             = _mm_sub_ps(iy3,jy1);
1757             dz31             = _mm_sub_ps(iz3,jz1);
1758             dx32             = _mm_sub_ps(ix3,jx2);
1759             dy32             = _mm_sub_ps(iy3,jy2);
1760             dz32             = _mm_sub_ps(iz3,jz2);
1761             dx33             = _mm_sub_ps(ix3,jx3);
1762             dy33             = _mm_sub_ps(iy3,jy3);
1763             dz33             = _mm_sub_ps(iz3,jz3);
1764
1765             /* Calculate squared distance and things based on it */
1766             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1767             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1768             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1769             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1770             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1771             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1772             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1773             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1774             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1775             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1776
1777             rinv11           = gmx_mm_invsqrt_ps(rsq11);
1778             rinv12           = gmx_mm_invsqrt_ps(rsq12);
1779             rinv13           = gmx_mm_invsqrt_ps(rsq13);
1780             rinv21           = gmx_mm_invsqrt_ps(rsq21);
1781             rinv22           = gmx_mm_invsqrt_ps(rsq22);
1782             rinv23           = gmx_mm_invsqrt_ps(rsq23);
1783             rinv31           = gmx_mm_invsqrt_ps(rsq31);
1784             rinv32           = gmx_mm_invsqrt_ps(rsq32);
1785             rinv33           = gmx_mm_invsqrt_ps(rsq33);
1786
1787             rinvsq00         = gmx_mm_inv_ps(rsq00);
1788             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1789             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1790             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1791             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1792             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1793             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1794             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1795             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1796             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1797
1798             fjx0             = _mm_setzero_ps();
1799             fjy0             = _mm_setzero_ps();
1800             fjz0             = _mm_setzero_ps();
1801             fjx1             = _mm_setzero_ps();
1802             fjy1             = _mm_setzero_ps();
1803             fjz1             = _mm_setzero_ps();
1804             fjx2             = _mm_setzero_ps();
1805             fjy2             = _mm_setzero_ps();
1806             fjz2             = _mm_setzero_ps();
1807             fjx3             = _mm_setzero_ps();
1808             fjy3             = _mm_setzero_ps();
1809             fjz3             = _mm_setzero_ps();
1810
1811             /**************************
1812              * CALCULATE INTERACTIONS *
1813              **************************/
1814
1815             if (gmx_mm_any_lt(rsq00,rcutoff2))
1816             {
1817
1818             /* LENNARD-JONES DISPERSION/REPULSION */
1819
1820             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1821             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1822
1823             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1824
1825             fscal            = fvdw;
1826
1827             fscal            = _mm_and_ps(fscal,cutoff_mask);
1828
1829             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1830
1831              /* Update vectorial force */
1832             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1833             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1834             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1835
1836             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1837             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1838             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1839
1840             }
1841
1842             /**************************
1843              * CALCULATE INTERACTIONS *
1844              **************************/
1845
1846             if (gmx_mm_any_lt(rsq11,rcutoff2))
1847             {
1848
1849             /* REACTION-FIELD ELECTROSTATICS */
1850             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1851
1852             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1853
1854             fscal            = felec;
1855
1856             fscal            = _mm_and_ps(fscal,cutoff_mask);
1857
1858             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1859
1860              /* Update vectorial force */
1861             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1862             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1863             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1864
1865             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1866             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1867             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1868
1869             }
1870
1871             /**************************
1872              * CALCULATE INTERACTIONS *
1873              **************************/
1874
1875             if (gmx_mm_any_lt(rsq12,rcutoff2))
1876             {
1877
1878             /* REACTION-FIELD ELECTROSTATICS */
1879             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1880
1881             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1882
1883             fscal            = felec;
1884
1885             fscal            = _mm_and_ps(fscal,cutoff_mask);
1886
1887             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1888
1889              /* Update vectorial force */
1890             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1891             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1892             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1893
1894             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1895             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1896             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1897
1898             }
1899
1900             /**************************
1901              * CALCULATE INTERACTIONS *
1902              **************************/
1903
1904             if (gmx_mm_any_lt(rsq13,rcutoff2))
1905             {
1906
1907             /* REACTION-FIELD ELECTROSTATICS */
1908             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
1909
1910             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
1911
1912             fscal            = felec;
1913
1914             fscal            = _mm_and_ps(fscal,cutoff_mask);
1915
1916             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1917
1918              /* Update vectorial force */
1919             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1920             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1921             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1922
1923             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1924             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1925             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1926
1927             }
1928
1929             /**************************
1930              * CALCULATE INTERACTIONS *
1931              **************************/
1932
1933             if (gmx_mm_any_lt(rsq21,rcutoff2))
1934             {
1935
1936             /* REACTION-FIELD ELECTROSTATICS */
1937             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1938
1939             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1940
1941             fscal            = felec;
1942
1943             fscal            = _mm_and_ps(fscal,cutoff_mask);
1944
1945             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1946
1947              /* Update vectorial force */
1948             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1949             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1950             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1951
1952             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1953             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1954             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1955
1956             }
1957
1958             /**************************
1959              * CALCULATE INTERACTIONS *
1960              **************************/
1961
1962             if (gmx_mm_any_lt(rsq22,rcutoff2))
1963             {
1964
1965             /* REACTION-FIELD ELECTROSTATICS */
1966             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1967
1968             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1969
1970             fscal            = felec;
1971
1972             fscal            = _mm_and_ps(fscal,cutoff_mask);
1973
1974             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1975
1976              /* Update vectorial force */
1977             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1978             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1979             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1980
1981             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1982             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1983             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1984
1985             }
1986
1987             /**************************
1988              * CALCULATE INTERACTIONS *
1989              **************************/
1990
1991             if (gmx_mm_any_lt(rsq23,rcutoff2))
1992             {
1993
1994             /* REACTION-FIELD ELECTROSTATICS */
1995             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
1996
1997             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
1998
1999             fscal            = felec;
2000
2001             fscal            = _mm_and_ps(fscal,cutoff_mask);
2002
2003             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2004
2005              /* Update vectorial force */
2006             fix2             = _mm_macc_ps(dx23,fscal,fix2);
2007             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
2008             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
2009
2010             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
2011             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
2012             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
2013
2014             }
2015
2016             /**************************
2017              * CALCULATE INTERACTIONS *
2018              **************************/
2019
2020             if (gmx_mm_any_lt(rsq31,rcutoff2))
2021             {
2022
2023             /* REACTION-FIELD ELECTROSTATICS */
2024             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
2025
2026             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
2027
2028             fscal            = felec;
2029
2030             fscal            = _mm_and_ps(fscal,cutoff_mask);
2031
2032             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2033
2034              /* Update vectorial force */
2035             fix3             = _mm_macc_ps(dx31,fscal,fix3);
2036             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
2037             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
2038
2039             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
2040             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
2041             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
2042
2043             }
2044
2045             /**************************
2046              * CALCULATE INTERACTIONS *
2047              **************************/
2048
2049             if (gmx_mm_any_lt(rsq32,rcutoff2))
2050             {
2051
2052             /* REACTION-FIELD ELECTROSTATICS */
2053             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
2054
2055             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
2056
2057             fscal            = felec;
2058
2059             fscal            = _mm_and_ps(fscal,cutoff_mask);
2060
2061             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2062
2063              /* Update vectorial force */
2064             fix3             = _mm_macc_ps(dx32,fscal,fix3);
2065             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
2066             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
2067
2068             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
2069             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
2070             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
2071
2072             }
2073
2074             /**************************
2075              * CALCULATE INTERACTIONS *
2076              **************************/
2077
2078             if (gmx_mm_any_lt(rsq33,rcutoff2))
2079             {
2080
2081             /* REACTION-FIELD ELECTROSTATICS */
2082             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
2083
2084             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
2085
2086             fscal            = felec;
2087
2088             fscal            = _mm_and_ps(fscal,cutoff_mask);
2089
2090             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2091
2092              /* Update vectorial force */
2093             fix3             = _mm_macc_ps(dx33,fscal,fix3);
2094             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
2095             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
2096
2097             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
2098             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
2099             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
2100
2101             }
2102
2103             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2104             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2105             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2106             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2107
2108             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
2109                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
2110                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2111
2112             /* Inner loop uses 333 flops */
2113         }
2114
2115         /* End of innermost loop */
2116
2117         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2118                                               f+i_coord_offset,fshift+i_shift_offset);
2119
2120         /* Increment number of inner iterations */
2121         inneriter                  += j_index_end - j_index_start;
2122
2123         /* Outer loop uses 24 flops */
2124     }
2125
2126     /* Increment number of outer iterations */
2127     outeriter        += nri;
2128
2129     /* Update outer/inner flops */
2130
2131     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*333);
2132 }