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

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017,2018, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_single kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_128_fma_single.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW4W4_VF_avx_128_fma_single
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Water4-Water4
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRFCut_VdwLJSh_GeomW4W4_VF_avx_128_fma_single
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB,jnrC,jnrD;
  74     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
  75     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
  80     real             scratch[4*DIM];
  81     __m128           fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwioffset1;
  85     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  86     int              vdwioffset2;
  87     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  88     int              vdwioffset3;
  89     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  90     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
  91     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  92     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
  93     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  94     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
  95     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  96     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
  97     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  98     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  99     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 100     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 101     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 102     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 103     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 104     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 105     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 106     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 107     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 108     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 109     real             *charge;
 110     int              nvdwtype;
 111     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 112     int              *vdwtype;
 113     real             *vdwparam;
 114     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
 115     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
 116     __m128           dummy_mask,cutoff_mask;
 117     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
 118     __m128           one     = _mm_set1_ps(1.0);
 119     __m128           two     = _mm_set1_ps(2.0);
 120     x                = xx[0];
 121     f                = ff[0];
 122
 123     nri              = nlist->nri;
 124     iinr             = nlist->iinr;
 125     jindex           = nlist->jindex;
 126     jjnr             = nlist->jjnr;
 127     shiftidx         = nlist->shift;
 128     gid              = nlist->gid;
 129     shiftvec         = fr->shift_vec[0];
 130     fshift           = fr->fshift[0];
 131     facel            = _mm_set1_ps(fr->ic->epsfac);
 132     charge           = mdatoms->chargeA;
 133     krf              = _mm_set1_ps(fr->ic->k_rf);
 134     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
 135     crf              = _mm_set1_ps(fr->ic->c_rf);
 136     nvdwtype         = fr->ntype;
 137     vdwparam         = fr->nbfp;
 138     vdwtype          = mdatoms->typeA;
 139
 140     /* Setup water-specific parameters */
 141     inr              = nlist->iinr[0];
 142     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 143     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 144     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 145     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 146
 147     jq1              = _mm_set1_ps(charge[inr+1]);
 148     jq2              = _mm_set1_ps(charge[inr+2]);
 149     jq3              = _mm_set1_ps(charge[inr+3]);
 150     vdwjidx0A        = 2*vdwtype[inr+0];
 151     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 152     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 153     qq11             = _mm_mul_ps(iq1,jq1);
 154     qq12             = _mm_mul_ps(iq1,jq2);
 155     qq13             = _mm_mul_ps(iq1,jq3);
 156     qq21             = _mm_mul_ps(iq2,jq1);
 157     qq22             = _mm_mul_ps(iq2,jq2);
 158     qq23             = _mm_mul_ps(iq2,jq3);
 159     qq31             = _mm_mul_ps(iq3,jq1);
 160     qq32             = _mm_mul_ps(iq3,jq2);
 161     qq33             = _mm_mul_ps(iq3,jq3);
 162
 163     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 164     rcutoff_scalar   = fr->ic->rcoulomb;
 165     rcutoff          = _mm_set1_ps(rcutoff_scalar);
 166     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
 167
 168     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
 169     rvdw             = _mm_set1_ps(fr->ic->rvdw);
 170
 171     /* Avoid stupid compiler warnings */
 172     jnrA = jnrB = jnrC = jnrD = 0;
 173     j_coord_offsetA = 0;
 174     j_coord_offsetB = 0;
 175     j_coord_offsetC = 0;
 176     j_coord_offsetD = 0;
 177
 178     outeriter        = 0;
 179     inneriter        = 0;
 180
 181     for(iidx=0;iidx<4*DIM;iidx++)
 182     {
 183         scratch[iidx] = 0.0;
 184     }
 185
 186     /* Start outer loop over neighborlists */
 187     for(iidx=0; iidx<nri; iidx++)
 188     {
 189         /* Load shift vector for this list */
 190         i_shift_offset   = DIM*shiftidx[iidx];
 191
 192         /* Load limits for loop over neighbors */
 193         j_index_start    = jindex[iidx];
 194         j_index_end      = jindex[iidx+1];
 195
 196         /* Get outer coordinate index */
 197         inr              = iinr[iidx];
 198         i_coord_offset   = DIM*inr;
 199
 200         /* Load i particle coords and add shift vector */
 201         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 202                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 203
 204         fix0             = _mm_setzero_ps();
 205         fiy0             = _mm_setzero_ps();
 206         fiz0             = _mm_setzero_ps();
 207         fix1             = _mm_setzero_ps();
 208         fiy1             = _mm_setzero_ps();
 209         fiz1             = _mm_setzero_ps();
 210         fix2             = _mm_setzero_ps();
 211         fiy2             = _mm_setzero_ps();
 212         fiz2             = _mm_setzero_ps();
 213         fix3             = _mm_setzero_ps();
 214         fiy3             = _mm_setzero_ps();
 215         fiz3             = _mm_setzero_ps();
 216
 217         /* Reset potential sums */
 218         velecsum         = _mm_setzero_ps();
 219         vvdwsum          = _mm_setzero_ps();
 220
 221         /* Start inner kernel loop */
 222         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 223         {
 224
 225             /* Get j neighbor index, and coordinate index */
 226             jnrA             = jjnr[jidx];
 227             jnrB             = jjnr[jidx+1];
 228             jnrC             = jjnr[jidx+2];
 229             jnrD             = jjnr[jidx+3];
 230             j_coord_offsetA  = DIM*jnrA;
 231             j_coord_offsetB  = DIM*jnrB;
 232             j_coord_offsetC  = DIM*jnrC;
 233             j_coord_offsetD  = DIM*jnrD;
 234
 235             /* load j atom coordinates */
 236             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 237                                               x+j_coord_offsetC,x+j_coord_offsetD,
 238                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 239                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 240
 241             /* Calculate displacement vector */
 242             dx00             = _mm_sub_ps(ix0,jx0);
 243             dy00             = _mm_sub_ps(iy0,jy0);
 244             dz00             = _mm_sub_ps(iz0,jz0);
 245             dx11             = _mm_sub_ps(ix1,jx1);
 246             dy11             = _mm_sub_ps(iy1,jy1);
 247             dz11             = _mm_sub_ps(iz1,jz1);
 248             dx12             = _mm_sub_ps(ix1,jx2);
 249             dy12             = _mm_sub_ps(iy1,jy2);
 250             dz12             = _mm_sub_ps(iz1,jz2);
 251             dx13             = _mm_sub_ps(ix1,jx3);
 252             dy13             = _mm_sub_ps(iy1,jy3);
 253             dz13             = _mm_sub_ps(iz1,jz3);
 254             dx21             = _mm_sub_ps(ix2,jx1);
 255             dy21             = _mm_sub_ps(iy2,jy1);
 256             dz21             = _mm_sub_ps(iz2,jz1);
 257             dx22             = _mm_sub_ps(ix2,jx2);
 258             dy22             = _mm_sub_ps(iy2,jy2);
 259             dz22             = _mm_sub_ps(iz2,jz2);
 260             dx23             = _mm_sub_ps(ix2,jx3);
 261             dy23             = _mm_sub_ps(iy2,jy3);
 262             dz23             = _mm_sub_ps(iz2,jz3);
 263             dx31             = _mm_sub_ps(ix3,jx1);
 264             dy31             = _mm_sub_ps(iy3,jy1);
 265             dz31             = _mm_sub_ps(iz3,jz1);
 266             dx32             = _mm_sub_ps(ix3,jx2);
 267             dy32             = _mm_sub_ps(iy3,jy2);
 268             dz32             = _mm_sub_ps(iz3,jz2);
 269             dx33             = _mm_sub_ps(ix3,jx3);
 270             dy33             = _mm_sub_ps(iy3,jy3);
 271             dz33             = _mm_sub_ps(iz3,jz3);
 272
 273             /* Calculate squared distance and things based on it */
 274             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 275             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 276             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 277             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 278             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 279             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 280             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 281             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 282             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 283             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 284
 285             rinv11           = avx128fma_invsqrt_f(rsq11);
 286             rinv12           = avx128fma_invsqrt_f(rsq12);
 287             rinv13           = avx128fma_invsqrt_f(rsq13);
 288             rinv21           = avx128fma_invsqrt_f(rsq21);
 289             rinv22           = avx128fma_invsqrt_f(rsq22);
 290             rinv23           = avx128fma_invsqrt_f(rsq23);
 291             rinv31           = avx128fma_invsqrt_f(rsq31);
 292             rinv32           = avx128fma_invsqrt_f(rsq32);
 293             rinv33           = avx128fma_invsqrt_f(rsq33);
 294
 295             rinvsq00         = avx128fma_inv_f(rsq00);
 296             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 297             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 298             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 299             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 300             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 301             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 302             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 303             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 304             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 305
 306             fjx0             = _mm_setzero_ps();
 307             fjy0             = _mm_setzero_ps();
 308             fjz0             = _mm_setzero_ps();
 309             fjx1             = _mm_setzero_ps();
 310             fjy1             = _mm_setzero_ps();
 311             fjz1             = _mm_setzero_ps();
 312             fjx2             = _mm_setzero_ps();
 313             fjy2             = _mm_setzero_ps();
 314             fjz2             = _mm_setzero_ps();
 315             fjx3             = _mm_setzero_ps();
 316             fjy3             = _mm_setzero_ps();
 317             fjz3             = _mm_setzero_ps();
 318
 319             /**************************
 320              * CALCULATE INTERACTIONS *
 321              **************************/
 322
 323             if (gmx_mm_any_lt(rsq00,rcutoff2))
 324             {
 325
 326             /* LENNARD-JONES DISPERSION/REPULSION */
 327
 328             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 329             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 330             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 331             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 332                                           _mm_mul_ps( _mm_nmacc_ps(c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 333             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 334
 335             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 336
 337             /* Update potential sum for this i atom from the interaction with this j atom. */
 338             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 339             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 340
 341             fscal            = fvdw;
 342
 343             fscal            = _mm_and_ps(fscal,cutoff_mask);
 344
 345              /* Update vectorial force */
 346             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 347             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 348             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 349
 350             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 351             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 352             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 353
 354             }
 355
 356             /**************************
 357              * CALCULATE INTERACTIONS *
 358              **************************/
 359
 360             if (gmx_mm_any_lt(rsq11,rcutoff2))
 361             {
 362
 363             /* REACTION-FIELD ELECTROSTATICS */
 364             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 365             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 366
 367             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 368
 369             /* Update potential sum for this i atom from the interaction with this j atom. */
 370             velec            = _mm_and_ps(velec,cutoff_mask);
 371             velecsum         = _mm_add_ps(velecsum,velec);
 372
 373             fscal            = felec;
 374
 375             fscal            = _mm_and_ps(fscal,cutoff_mask);
 376
 377              /* Update vectorial force */
 378             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 379             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 380             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 381
 382             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 383             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 384             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 385
 386             }
 387
 388             /**************************
 389              * CALCULATE INTERACTIONS *
 390              **************************/
 391
 392             if (gmx_mm_any_lt(rsq12,rcutoff2))
 393             {
 394
 395             /* REACTION-FIELD ELECTROSTATICS */
 396             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 397             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 398
 399             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 400
 401             /* Update potential sum for this i atom from the interaction with this j atom. */
 402             velec            = _mm_and_ps(velec,cutoff_mask);
 403             velecsum         = _mm_add_ps(velecsum,velec);
 404
 405             fscal            = felec;
 406
 407             fscal            = _mm_and_ps(fscal,cutoff_mask);
 408
 409              /* Update vectorial force */
 410             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 411             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 412             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 413
 414             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 415             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 416             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 417
 418             }
 419
 420             /**************************
 421              * CALCULATE INTERACTIONS *
 422              **************************/
 423
 424             if (gmx_mm_any_lt(rsq13,rcutoff2))
 425             {
 426
 427             /* REACTION-FIELD ELECTROSTATICS */
 428             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_macc_ps(krf,rsq13,rinv13),crf));
 429             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
 430
 431             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
 432
 433             /* Update potential sum for this i atom from the interaction with this j atom. */
 434             velec            = _mm_and_ps(velec,cutoff_mask);
 435             velecsum         = _mm_add_ps(velecsum,velec);
 436
 437             fscal            = felec;
 438
 439             fscal            = _mm_and_ps(fscal,cutoff_mask);
 440
 441              /* Update vectorial force */
 442             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 443             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 444             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 445
 446             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 447             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 448             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 449
 450             }
 451
 452             /**************************
 453              * CALCULATE INTERACTIONS *
 454              **************************/
 455
 456             if (gmx_mm_any_lt(rsq21,rcutoff2))
 457             {
 458
 459             /* REACTION-FIELD ELECTROSTATICS */
 460             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 461             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 462
 463             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 464
 465             /* Update potential sum for this i atom from the interaction with this j atom. */
 466             velec            = _mm_and_ps(velec,cutoff_mask);
 467             velecsum         = _mm_add_ps(velecsum,velec);
 468
 469             fscal            = felec;
 470
 471             fscal            = _mm_and_ps(fscal,cutoff_mask);
 472
 473              /* Update vectorial force */
 474             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 475             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 476             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 477
 478             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 479             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 480             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 481
 482             }
 483
 484             /**************************
 485              * CALCULATE INTERACTIONS *
 486              **************************/
 487
 488             if (gmx_mm_any_lt(rsq22,rcutoff2))
 489             {
 490
 491             /* REACTION-FIELD ELECTROSTATICS */
 492             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 493             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 494
 495             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 496
 497             /* Update potential sum for this i atom from the interaction with this j atom. */
 498             velec            = _mm_and_ps(velec,cutoff_mask);
 499             velecsum         = _mm_add_ps(velecsum,velec);
 500
 501             fscal            = felec;
 502
 503             fscal            = _mm_and_ps(fscal,cutoff_mask);
 504
 505              /* Update vectorial force */
 506             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 507             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 508             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 509
 510             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 511             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 512             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 513
 514             }
 515
 516             /**************************
 517              * CALCULATE INTERACTIONS *
 518              **************************/
 519
 520             if (gmx_mm_any_lt(rsq23,rcutoff2))
 521             {
 522
 523             /* REACTION-FIELD ELECTROSTATICS */
 524             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_macc_ps(krf,rsq23,rinv23),crf));
 525             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
 526
 527             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
 528
 529             /* Update potential sum for this i atom from the interaction with this j atom. */
 530             velec            = _mm_and_ps(velec,cutoff_mask);
 531             velecsum         = _mm_add_ps(velecsum,velec);
 532
 533             fscal            = felec;
 534
 535             fscal            = _mm_and_ps(fscal,cutoff_mask);
 536
 537              /* Update vectorial force */
 538             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 539             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 540             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 541
 542             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 543             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 544             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 545
 546             }
 547
 548             /**************************
 549              * CALCULATE INTERACTIONS *
 550              **************************/
 551
 552             if (gmx_mm_any_lt(rsq31,rcutoff2))
 553             {
 554
 555             /* REACTION-FIELD ELECTROSTATICS */
 556             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_macc_ps(krf,rsq31,rinv31),crf));
 557             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
 558
 559             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
 560
 561             /* Update potential sum for this i atom from the interaction with this j atom. */
 562             velec            = _mm_and_ps(velec,cutoff_mask);
 563             velecsum         = _mm_add_ps(velecsum,velec);
 564
 565             fscal            = felec;
 566
 567             fscal            = _mm_and_ps(fscal,cutoff_mask);
 568
 569              /* Update vectorial force */
 570             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 571             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 572             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 573
 574             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 575             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 576             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 577
 578             }
 579
 580             /**************************
 581              * CALCULATE INTERACTIONS *
 582              **************************/
 583
 584             if (gmx_mm_any_lt(rsq32,rcutoff2))
 585             {
 586
 587             /* REACTION-FIELD ELECTROSTATICS */
 588             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_macc_ps(krf,rsq32,rinv32),crf));
 589             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
 590
 591             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
 592
 593             /* Update potential sum for this i atom from the interaction with this j atom. */
 594             velec            = _mm_and_ps(velec,cutoff_mask);
 595             velecsum         = _mm_add_ps(velecsum,velec);
 596
 597             fscal            = felec;
 598
 599             fscal            = _mm_and_ps(fscal,cutoff_mask);
 600
 601              /* Update vectorial force */
 602             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 603             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 604             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 605
 606             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 607             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 608             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 609
 610             }
 611
 612             /**************************
 613              * CALCULATE INTERACTIONS *
 614              **************************/
 615
 616             if (gmx_mm_any_lt(rsq33,rcutoff2))
 617             {
 618
 619             /* REACTION-FIELD ELECTROSTATICS */
 620             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_macc_ps(krf,rsq33,rinv33),crf));
 621             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
 622
 623             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
 624
 625             /* Update potential sum for this i atom from the interaction with this j atom. */
 626             velec            = _mm_and_ps(velec,cutoff_mask);
 627             velecsum         = _mm_add_ps(velecsum,velec);
 628
 629             fscal            = felec;
 630
 631             fscal            = _mm_and_ps(fscal,cutoff_mask);
 632
 633              /* Update vectorial force */
 634             fix3             = _mm_macc_ps(dx33,fscal,fix3);
 635             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
 636             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
 637
 638             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
 639             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
 640             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
 641
 642             }
 643
 644             fjptrA             = f+j_coord_offsetA;
 645             fjptrB             = f+j_coord_offsetB;
 646             fjptrC             = f+j_coord_offsetC;
 647             fjptrD             = f+j_coord_offsetD;
 648
 649             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 650                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 651                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 652
 653             /* Inner loop uses 398 flops */
 654         }
 655
 656         if(jidx<j_index_end)
 657         {
 658
 659             /* Get j neighbor index, and coordinate index */
 660             jnrlistA         = jjnr[jidx];
 661             jnrlistB         = jjnr[jidx+1];
 662             jnrlistC         = jjnr[jidx+2];
 663             jnrlistD         = jjnr[jidx+3];
 664             /* Sign of each element will be negative for non-real atoms.
 665              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 666              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 667              */
 668             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 669             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 670             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 671             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 672             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 673             j_coord_offsetA  = DIM*jnrA;
 674             j_coord_offsetB  = DIM*jnrB;
 675             j_coord_offsetC  = DIM*jnrC;
 676             j_coord_offsetD  = DIM*jnrD;
 677
 678             /* load j atom coordinates */
 679             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 680                                               x+j_coord_offsetC,x+j_coord_offsetD,
 681                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 682                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 683
 684             /* Calculate displacement vector */
 685             dx00             = _mm_sub_ps(ix0,jx0);
 686             dy00             = _mm_sub_ps(iy0,jy0);
 687             dz00             = _mm_sub_ps(iz0,jz0);
 688             dx11             = _mm_sub_ps(ix1,jx1);
 689             dy11             = _mm_sub_ps(iy1,jy1);
 690             dz11             = _mm_sub_ps(iz1,jz1);
 691             dx12             = _mm_sub_ps(ix1,jx2);
 692             dy12             = _mm_sub_ps(iy1,jy2);
 693             dz12             = _mm_sub_ps(iz1,jz2);
 694             dx13             = _mm_sub_ps(ix1,jx3);
 695             dy13             = _mm_sub_ps(iy1,jy3);
 696             dz13             = _mm_sub_ps(iz1,jz3);
 697             dx21             = _mm_sub_ps(ix2,jx1);
 698             dy21             = _mm_sub_ps(iy2,jy1);
 699             dz21             = _mm_sub_ps(iz2,jz1);
 700             dx22             = _mm_sub_ps(ix2,jx2);
 701             dy22             = _mm_sub_ps(iy2,jy2);
 702             dz22             = _mm_sub_ps(iz2,jz2);
 703             dx23             = _mm_sub_ps(ix2,jx3);
 704             dy23             = _mm_sub_ps(iy2,jy3);
 705             dz23             = _mm_sub_ps(iz2,jz3);
 706             dx31             = _mm_sub_ps(ix3,jx1);
 707             dy31             = _mm_sub_ps(iy3,jy1);
 708             dz31             = _mm_sub_ps(iz3,jz1);
 709             dx32             = _mm_sub_ps(ix3,jx2);
 710             dy32             = _mm_sub_ps(iy3,jy2);
 711             dz32             = _mm_sub_ps(iz3,jz2);
 712             dx33             = _mm_sub_ps(ix3,jx3);
 713             dy33             = _mm_sub_ps(iy3,jy3);
 714             dz33             = _mm_sub_ps(iz3,jz3);
 715
 716             /* Calculate squared distance and things based on it */
 717             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 718             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 719             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 720             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 721             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 722             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 723             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 724             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 725             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 726             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 727
 728             rinv11           = avx128fma_invsqrt_f(rsq11);
 729             rinv12           = avx128fma_invsqrt_f(rsq12);
 730             rinv13           = avx128fma_invsqrt_f(rsq13);
 731             rinv21           = avx128fma_invsqrt_f(rsq21);
 732             rinv22           = avx128fma_invsqrt_f(rsq22);
 733             rinv23           = avx128fma_invsqrt_f(rsq23);
 734             rinv31           = avx128fma_invsqrt_f(rsq31);
 735             rinv32           = avx128fma_invsqrt_f(rsq32);
 736             rinv33           = avx128fma_invsqrt_f(rsq33);
 737
 738             rinvsq00         = avx128fma_inv_f(rsq00);
 739             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 740             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 741             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 742             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 743             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 744             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 745             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 746             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 747             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 748
 749             fjx0             = _mm_setzero_ps();
 750             fjy0             = _mm_setzero_ps();
 751             fjz0             = _mm_setzero_ps();
 752             fjx1             = _mm_setzero_ps();
 753             fjy1             = _mm_setzero_ps();
 754             fjz1             = _mm_setzero_ps();
 755             fjx2             = _mm_setzero_ps();
 756             fjy2             = _mm_setzero_ps();
 757             fjz2             = _mm_setzero_ps();
 758             fjx3             = _mm_setzero_ps();
 759             fjy3             = _mm_setzero_ps();
 760             fjz3             = _mm_setzero_ps();
 761
 762             /**************************
 763              * CALCULATE INTERACTIONS *
 764              **************************/
 765
 766             if (gmx_mm_any_lt(rsq00,rcutoff2))
 767             {
 768
 769             /* LENNARD-JONES DISPERSION/REPULSION */
 770
 771             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 772             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 773             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 774             vvdw             = _mm_msub_ps(_mm_nmacc_ps(c12_00,_mm_mul_ps(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 775                                           _mm_mul_ps( _mm_nmacc_ps(c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 776             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 777
 778             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
 779
 780             /* Update potential sum for this i atom from the interaction with this j atom. */
 781             vvdw             = _mm_and_ps(vvdw,cutoff_mask);
 782             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 783             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 784
 785             fscal            = fvdw;
 786
 787             fscal            = _mm_and_ps(fscal,cutoff_mask);
 788
 789             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 790
 791              /* Update vectorial force */
 792             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 793             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 794             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 795
 796             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 797             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 798             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 799
 800             }
 801
 802             /**************************
 803              * CALCULATE INTERACTIONS *
 804              **************************/
 805
 806             if (gmx_mm_any_lt(rsq11,rcutoff2))
 807             {
 808
 809             /* REACTION-FIELD ELECTROSTATICS */
 810             velec            = _mm_mul_ps(qq11,_mm_sub_ps(_mm_macc_ps(krf,rsq11,rinv11),crf));
 811             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
 812
 813             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
 814
 815             /* Update potential sum for this i atom from the interaction with this j atom. */
 816             velec            = _mm_and_ps(velec,cutoff_mask);
 817             velec            = _mm_andnot_ps(dummy_mask,velec);
 818             velecsum         = _mm_add_ps(velecsum,velec);
 819
 820             fscal            = felec;
 821
 822             fscal            = _mm_and_ps(fscal,cutoff_mask);
 823
 824             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 825
 826              /* Update vectorial force */
 827             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 828             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 829             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 830
 831             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 832             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 833             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 834
 835             }
 836
 837             /**************************
 838              * CALCULATE INTERACTIONS *
 839              **************************/
 840
 841             if (gmx_mm_any_lt(rsq12,rcutoff2))
 842             {
 843
 844             /* REACTION-FIELD ELECTROSTATICS */
 845             velec            = _mm_mul_ps(qq12,_mm_sub_ps(_mm_macc_ps(krf,rsq12,rinv12),crf));
 846             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
 847
 848             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
 849
 850             /* Update potential sum for this i atom from the interaction with this j atom. */
 851             velec            = _mm_and_ps(velec,cutoff_mask);
 852             velec            = _mm_andnot_ps(dummy_mask,velec);
 853             velecsum         = _mm_add_ps(velecsum,velec);
 854
 855             fscal            = felec;
 856
 857             fscal            = _mm_and_ps(fscal,cutoff_mask);
 858
 859             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 860
 861              /* Update vectorial force */
 862             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 863             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 864             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 865
 866             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 867             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 868             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 869
 870             }
 871
 872             /**************************
 873              * CALCULATE INTERACTIONS *
 874              **************************/
 875
 876             if (gmx_mm_any_lt(rsq13,rcutoff2))
 877             {
 878
 879             /* REACTION-FIELD ELECTROSTATICS */
 880             velec            = _mm_mul_ps(qq13,_mm_sub_ps(_mm_macc_ps(krf,rsq13,rinv13),crf));
 881             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
 882
 883             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
 884
 885             /* Update potential sum for this i atom from the interaction with this j atom. */
 886             velec            = _mm_and_ps(velec,cutoff_mask);
 887             velec            = _mm_andnot_ps(dummy_mask,velec);
 888             velecsum         = _mm_add_ps(velecsum,velec);
 889
 890             fscal            = felec;
 891
 892             fscal            = _mm_and_ps(fscal,cutoff_mask);
 893
 894             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 895
 896              /* Update vectorial force */
 897             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 898             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 899             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 900
 901             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 902             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 903             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 904
 905             }
 906
 907             /**************************
 908              * CALCULATE INTERACTIONS *
 909              **************************/
 910
 911             if (gmx_mm_any_lt(rsq21,rcutoff2))
 912             {
 913
 914             /* REACTION-FIELD ELECTROSTATICS */
 915             velec            = _mm_mul_ps(qq21,_mm_sub_ps(_mm_macc_ps(krf,rsq21,rinv21),crf));
 916             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
 917
 918             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
 919
 920             /* Update potential sum for this i atom from the interaction with this j atom. */
 921             velec            = _mm_and_ps(velec,cutoff_mask);
 922             velec            = _mm_andnot_ps(dummy_mask,velec);
 923             velecsum         = _mm_add_ps(velecsum,velec);
 924
 925             fscal            = felec;
 926
 927             fscal            = _mm_and_ps(fscal,cutoff_mask);
 928
 929             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 930
 931              /* Update vectorial force */
 932             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 933             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 934             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 935
 936             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 937             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 938             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 939
 940             }
 941
 942             /**************************
 943              * CALCULATE INTERACTIONS *
 944              **************************/
 945
 946             if (gmx_mm_any_lt(rsq22,rcutoff2))
 947             {
 948
 949             /* REACTION-FIELD ELECTROSTATICS */
 950             velec            = _mm_mul_ps(qq22,_mm_sub_ps(_mm_macc_ps(krf,rsq22,rinv22),crf));
 951             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
 952
 953             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
 954
 955             /* Update potential sum for this i atom from the interaction with this j atom. */
 956             velec            = _mm_and_ps(velec,cutoff_mask);
 957             velec            = _mm_andnot_ps(dummy_mask,velec);
 958             velecsum         = _mm_add_ps(velecsum,velec);
 959
 960             fscal            = felec;
 961
 962             fscal            = _mm_and_ps(fscal,cutoff_mask);
 963
 964             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 965
 966              /* Update vectorial force */
 967             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 968             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 969             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 970
 971             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 972             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 973             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 974
 975             }
 976
 977             /**************************
 978              * CALCULATE INTERACTIONS *
 979              **************************/
 980
 981             if (gmx_mm_any_lt(rsq23,rcutoff2))
 982             {
 983
 984             /* REACTION-FIELD ELECTROSTATICS */
 985             velec            = _mm_mul_ps(qq23,_mm_sub_ps(_mm_macc_ps(krf,rsq23,rinv23),crf));
 986             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
 987
 988             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
 989
 990             /* Update potential sum for this i atom from the interaction with this j atom. */
 991             velec            = _mm_and_ps(velec,cutoff_mask);
 992             velec            = _mm_andnot_ps(dummy_mask,velec);
 993             velecsum         = _mm_add_ps(velecsum,velec);
 994
 995             fscal            = felec;
 996
 997             fscal            = _mm_and_ps(fscal,cutoff_mask);
 998
 999             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1000
1001              /* Update vectorial force */
1002             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1003             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1004             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1005
1006             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1007             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1008             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1009
1010             }
1011
1012             /**************************
1013              * CALCULATE INTERACTIONS *
1014              **************************/
1015
1016             if (gmx_mm_any_lt(rsq31,rcutoff2))
1017             {
1018
1019             /* REACTION-FIELD ELECTROSTATICS */
1020             velec            = _mm_mul_ps(qq31,_mm_sub_ps(_mm_macc_ps(krf,rsq31,rinv31),crf));
1021             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
1022
1023             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
1024
1025             /* Update potential sum for this i atom from the interaction with this j atom. */
1026             velec            = _mm_and_ps(velec,cutoff_mask);
1027             velec            = _mm_andnot_ps(dummy_mask,velec);
1028             velecsum         = _mm_add_ps(velecsum,velec);
1029
1030             fscal            = felec;
1031
1032             fscal            = _mm_and_ps(fscal,cutoff_mask);
1033
1034             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1035
1036              /* Update vectorial force */
1037             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1038             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1039             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1040
1041             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1042             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1043             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1044
1045             }
1046
1047             /**************************
1048              * CALCULATE INTERACTIONS *
1049              **************************/
1050
1051             if (gmx_mm_any_lt(rsq32,rcutoff2))
1052             {
1053
1054             /* REACTION-FIELD ELECTROSTATICS */
1055             velec            = _mm_mul_ps(qq32,_mm_sub_ps(_mm_macc_ps(krf,rsq32,rinv32),crf));
1056             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
1057
1058             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1059
1060             /* Update potential sum for this i atom from the interaction with this j atom. */
1061             velec            = _mm_and_ps(velec,cutoff_mask);
1062             velec            = _mm_andnot_ps(dummy_mask,velec);
1063             velecsum         = _mm_add_ps(velecsum,velec);
1064
1065             fscal            = felec;
1066
1067             fscal            = _mm_and_ps(fscal,cutoff_mask);
1068
1069             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1070
1071              /* Update vectorial force */
1072             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1073             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1074             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1075
1076             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1077             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1078             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1079
1080             }
1081
1082             /**************************
1083              * CALCULATE INTERACTIONS *
1084              **************************/
1085
1086             if (gmx_mm_any_lt(rsq33,rcutoff2))
1087             {
1088
1089             /* REACTION-FIELD ELECTROSTATICS */
1090             velec            = _mm_mul_ps(qq33,_mm_sub_ps(_mm_macc_ps(krf,rsq33,rinv33),crf));
1091             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
1092
1093             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
1094
1095             /* Update potential sum for this i atom from the interaction with this j atom. */
1096             velec            = _mm_and_ps(velec,cutoff_mask);
1097             velec            = _mm_andnot_ps(dummy_mask,velec);
1098             velecsum         = _mm_add_ps(velecsum,velec);
1099
1100             fscal            = felec;
1101
1102             fscal            = _mm_and_ps(fscal,cutoff_mask);
1103
1104             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1105
1106              /* Update vectorial force */
1107             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1108             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1109             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1110
1111             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1112             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1113             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1114
1115             }
1116
1117             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1118             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1119             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1120             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1121
1122             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1123                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1124                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1125
1126             /* Inner loop uses 398 flops */
1127         }
1128
1129         /* End of innermost loop */
1130
1131         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1132                                               f+i_coord_offset,fshift+i_shift_offset);
1133
1134         ggid                        = gid[iidx];
1135         /* Update potential energies */
1136         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
1137         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
1138
1139         /* Increment number of inner iterations */
1140         inneriter                  += j_index_end - j_index_start;
1141
1142         /* Outer loop uses 26 flops */
1143     }
1144
1145     /* Increment number of outer iterations */
1146     outeriter        += nri;
1147
1148     /* Update outer/inner flops */
1149
1150     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*398);
1151 }
1152 /*
1153  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW4W4_F_avx_128_fma_single
1154  * Electrostatics interaction: ReactionField
1155  * VdW interaction:            LennardJones
1156  * Geometry:                   Water4-Water4
1157  * Calculate force/pot:        Force
1158  */
1159 void
1160 nb_kernel_ElecRFCut_VdwLJSh_GeomW4W4_F_avx_128_fma_single
1161                     (t_nblist                    * gmx_restrict       nlist,
1162                      rvec                        * gmx_restrict          xx,
1163                      rvec                        * gmx_restrict          ff,
1164                      struct t_forcerec           * gmx_restrict          fr,
1165                      t_mdatoms                   * gmx_restrict     mdatoms,
1166                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1167                      t_nrnb                      * gmx_restrict        nrnb)
1168 {
1169     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1170      * just 0 for non-waters.
1171      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
1172      * jnr indices corresponding to data put in the four positions in the SIMD register.
1173      */
1174     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1175     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1176     int              jnrA,jnrB,jnrC,jnrD;
1177     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
1178     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
1179     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1180     real             rcutoff_scalar;
1181     real             *shiftvec,*fshift,*x,*f;
1182     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
1183     real             scratch[4*DIM];
1184     __m128           fscal,rcutoff,rcutoff2,jidxall;
1185     int              vdwioffset0;
1186     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1187     int              vdwioffset1;
1188     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1189     int              vdwioffset2;
1190     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1191     int              vdwioffset3;
1192     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
1193     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
1194     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1195     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
1196     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1197     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
1198     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1199     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
1200     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
1201     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1202     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1203     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1204     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
1205     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1206     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1207     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
1208     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
1209     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
1210     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
1211     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
1212     real             *charge;
1213     int              nvdwtype;
1214     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1215     int              *vdwtype;
1216     real             *vdwparam;
1217     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1218     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1219     __m128           dummy_mask,cutoff_mask;
1220     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1221     __m128           one     = _mm_set1_ps(1.0);
1222     __m128           two     = _mm_set1_ps(2.0);
1223     x                = xx[0];
1224     f                = ff[0];
1225
1226     nri              = nlist->nri;
1227     iinr             = nlist->iinr;
1228     jindex           = nlist->jindex;
1229     jjnr             = nlist->jjnr;
1230     shiftidx         = nlist->shift;
1231     gid              = nlist->gid;
1232     shiftvec         = fr->shift_vec[0];
1233     fshift           = fr->fshift[0];
1234     facel            = _mm_set1_ps(fr->ic->epsfac);
1235     charge           = mdatoms->chargeA;
1236     krf              = _mm_set1_ps(fr->ic->k_rf);
1237     krf2             = _mm_set1_ps(fr->ic->k_rf*2.0);
1238     crf              = _mm_set1_ps(fr->ic->c_rf);
1239     nvdwtype         = fr->ntype;
1240     vdwparam         = fr->nbfp;
1241     vdwtype          = mdatoms->typeA;
1242
1243     /* Setup water-specific parameters */
1244     inr              = nlist->iinr[0];
1245     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1246     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1247     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1248     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1249
1250     jq1              = _mm_set1_ps(charge[inr+1]);
1251     jq2              = _mm_set1_ps(charge[inr+2]);
1252     jq3              = _mm_set1_ps(charge[inr+3]);
1253     vdwjidx0A        = 2*vdwtype[inr+0];
1254     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1255     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1256     qq11             = _mm_mul_ps(iq1,jq1);
1257     qq12             = _mm_mul_ps(iq1,jq2);
1258     qq13             = _mm_mul_ps(iq1,jq3);
1259     qq21             = _mm_mul_ps(iq2,jq1);
1260     qq22             = _mm_mul_ps(iq2,jq2);
1261     qq23             = _mm_mul_ps(iq2,jq3);
1262     qq31             = _mm_mul_ps(iq3,jq1);
1263     qq32             = _mm_mul_ps(iq3,jq2);
1264     qq33             = _mm_mul_ps(iq3,jq3);
1265
1266     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1267     rcutoff_scalar   = fr->ic->rcoulomb;
1268     rcutoff          = _mm_set1_ps(rcutoff_scalar);
1269     rcutoff2         = _mm_mul_ps(rcutoff,rcutoff);
1270
1271     sh_vdw_invrcut6  = _mm_set1_ps(fr->ic->sh_invrc6);
1272     rvdw             = _mm_set1_ps(fr->ic->rvdw);
1273
1274     /* Avoid stupid compiler warnings */
1275     jnrA = jnrB = jnrC = jnrD = 0;
1276     j_coord_offsetA = 0;
1277     j_coord_offsetB = 0;
1278     j_coord_offsetC = 0;
1279     j_coord_offsetD = 0;
1280
1281     outeriter        = 0;
1282     inneriter        = 0;
1283
1284     for(iidx=0;iidx<4*DIM;iidx++)
1285     {
1286         scratch[iidx] = 0.0;
1287     }
1288
1289     /* Start outer loop over neighborlists */
1290     for(iidx=0; iidx<nri; iidx++)
1291     {
1292         /* Load shift vector for this list */
1293         i_shift_offset   = DIM*shiftidx[iidx];
1294
1295         /* Load limits for loop over neighbors */
1296         j_index_start    = jindex[iidx];
1297         j_index_end      = jindex[iidx+1];
1298
1299         /* Get outer coordinate index */
1300         inr              = iinr[iidx];
1301         i_coord_offset   = DIM*inr;
1302
1303         /* Load i particle coords and add shift vector */
1304         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1305                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1306
1307         fix0             = _mm_setzero_ps();
1308         fiy0             = _mm_setzero_ps();
1309         fiz0             = _mm_setzero_ps();
1310         fix1             = _mm_setzero_ps();
1311         fiy1             = _mm_setzero_ps();
1312         fiz1             = _mm_setzero_ps();
1313         fix2             = _mm_setzero_ps();
1314         fiy2             = _mm_setzero_ps();
1315         fiz2             = _mm_setzero_ps();
1316         fix3             = _mm_setzero_ps();
1317         fiy3             = _mm_setzero_ps();
1318         fiz3             = _mm_setzero_ps();
1319
1320         /* Start inner kernel loop */
1321         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1322         {
1323
1324             /* Get j neighbor index, and coordinate index */
1325             jnrA             = jjnr[jidx];
1326             jnrB             = jjnr[jidx+1];
1327             jnrC             = jjnr[jidx+2];
1328             jnrD             = jjnr[jidx+3];
1329             j_coord_offsetA  = DIM*jnrA;
1330             j_coord_offsetB  = DIM*jnrB;
1331             j_coord_offsetC  = DIM*jnrC;
1332             j_coord_offsetD  = DIM*jnrD;
1333
1334             /* load j atom coordinates */
1335             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1336                                               x+j_coord_offsetC,x+j_coord_offsetD,
1337                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1338                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1339
1340             /* Calculate displacement vector */
1341             dx00             = _mm_sub_ps(ix0,jx0);
1342             dy00             = _mm_sub_ps(iy0,jy0);
1343             dz00             = _mm_sub_ps(iz0,jz0);
1344             dx11             = _mm_sub_ps(ix1,jx1);
1345             dy11             = _mm_sub_ps(iy1,jy1);
1346             dz11             = _mm_sub_ps(iz1,jz1);
1347             dx12             = _mm_sub_ps(ix1,jx2);
1348             dy12             = _mm_sub_ps(iy1,jy2);
1349             dz12             = _mm_sub_ps(iz1,jz2);
1350             dx13             = _mm_sub_ps(ix1,jx3);
1351             dy13             = _mm_sub_ps(iy1,jy3);
1352             dz13             = _mm_sub_ps(iz1,jz3);
1353             dx21             = _mm_sub_ps(ix2,jx1);
1354             dy21             = _mm_sub_ps(iy2,jy1);
1355             dz21             = _mm_sub_ps(iz2,jz1);
1356             dx22             = _mm_sub_ps(ix2,jx2);
1357             dy22             = _mm_sub_ps(iy2,jy2);
1358             dz22             = _mm_sub_ps(iz2,jz2);
1359             dx23             = _mm_sub_ps(ix2,jx3);
1360             dy23             = _mm_sub_ps(iy2,jy3);
1361             dz23             = _mm_sub_ps(iz2,jz3);
1362             dx31             = _mm_sub_ps(ix3,jx1);
1363             dy31             = _mm_sub_ps(iy3,jy1);
1364             dz31             = _mm_sub_ps(iz3,jz1);
1365             dx32             = _mm_sub_ps(ix3,jx2);
1366             dy32             = _mm_sub_ps(iy3,jy2);
1367             dz32             = _mm_sub_ps(iz3,jz2);
1368             dx33             = _mm_sub_ps(ix3,jx3);
1369             dy33             = _mm_sub_ps(iy3,jy3);
1370             dz33             = _mm_sub_ps(iz3,jz3);
1371
1372             /* Calculate squared distance and things based on it */
1373             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1374             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1375             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1376             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1377             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1378             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1379             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1380             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1381             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1382             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1383
1384             rinv11           = avx128fma_invsqrt_f(rsq11);
1385             rinv12           = avx128fma_invsqrt_f(rsq12);
1386             rinv13           = avx128fma_invsqrt_f(rsq13);
1387             rinv21           = avx128fma_invsqrt_f(rsq21);
1388             rinv22           = avx128fma_invsqrt_f(rsq22);
1389             rinv23           = avx128fma_invsqrt_f(rsq23);
1390             rinv31           = avx128fma_invsqrt_f(rsq31);
1391             rinv32           = avx128fma_invsqrt_f(rsq32);
1392             rinv33           = avx128fma_invsqrt_f(rsq33);
1393
1394             rinvsq00         = avx128fma_inv_f(rsq00);
1395             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1396             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1397             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1398             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1399             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1400             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1401             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1402             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1403             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1404
1405             fjx0             = _mm_setzero_ps();
1406             fjy0             = _mm_setzero_ps();
1407             fjz0             = _mm_setzero_ps();
1408             fjx1             = _mm_setzero_ps();
1409             fjy1             = _mm_setzero_ps();
1410             fjz1             = _mm_setzero_ps();
1411             fjx2             = _mm_setzero_ps();
1412             fjy2             = _mm_setzero_ps();
1413             fjz2             = _mm_setzero_ps();
1414             fjx3             = _mm_setzero_ps();
1415             fjy3             = _mm_setzero_ps();
1416             fjz3             = _mm_setzero_ps();
1417
1418             /**************************
1419              * CALCULATE INTERACTIONS *
1420              **************************/
1421
1422             if (gmx_mm_any_lt(rsq00,rcutoff2))
1423             {
1424
1425             /* LENNARD-JONES DISPERSION/REPULSION */
1426
1427             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1428             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1429
1430             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1431
1432             fscal            = fvdw;
1433
1434             fscal            = _mm_and_ps(fscal,cutoff_mask);
1435
1436              /* Update vectorial force */
1437             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1438             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1439             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1440
1441             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1442             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1443             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1444
1445             }
1446
1447             /**************************
1448              * CALCULATE INTERACTIONS *
1449              **************************/
1450
1451             if (gmx_mm_any_lt(rsq11,rcutoff2))
1452             {
1453
1454             /* REACTION-FIELD ELECTROSTATICS */
1455             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1456
1457             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1458
1459             fscal            = felec;
1460
1461             fscal            = _mm_and_ps(fscal,cutoff_mask);
1462
1463              /* Update vectorial force */
1464             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1465             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1466             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1467
1468             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1469             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1470             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1471
1472             }
1473
1474             /**************************
1475              * CALCULATE INTERACTIONS *
1476              **************************/
1477
1478             if (gmx_mm_any_lt(rsq12,rcutoff2))
1479             {
1480
1481             /* REACTION-FIELD ELECTROSTATICS */
1482             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1483
1484             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1485
1486             fscal            = felec;
1487
1488             fscal            = _mm_and_ps(fscal,cutoff_mask);
1489
1490              /* Update vectorial force */
1491             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1492             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1493             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1494
1495             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1496             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1497             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1498
1499             }
1500
1501             /**************************
1502              * CALCULATE INTERACTIONS *
1503              **************************/
1504
1505             if (gmx_mm_any_lt(rsq13,rcutoff2))
1506             {
1507
1508             /* REACTION-FIELD ELECTROSTATICS */
1509             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
1510
1511             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
1512
1513             fscal            = felec;
1514
1515             fscal            = _mm_and_ps(fscal,cutoff_mask);
1516
1517              /* Update vectorial force */
1518             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1519             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1520             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1521
1522             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1523             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1524             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1525
1526             }
1527
1528             /**************************
1529              * CALCULATE INTERACTIONS *
1530              **************************/
1531
1532             if (gmx_mm_any_lt(rsq21,rcutoff2))
1533             {
1534
1535             /* REACTION-FIELD ELECTROSTATICS */
1536             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1537
1538             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1539
1540             fscal            = felec;
1541
1542             fscal            = _mm_and_ps(fscal,cutoff_mask);
1543
1544              /* Update vectorial force */
1545             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1546             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1547             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1548
1549             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1550             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1551             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1552
1553             }
1554
1555             /**************************
1556              * CALCULATE INTERACTIONS *
1557              **************************/
1558
1559             if (gmx_mm_any_lt(rsq22,rcutoff2))
1560             {
1561
1562             /* REACTION-FIELD ELECTROSTATICS */
1563             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1564
1565             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1566
1567             fscal            = felec;
1568
1569             fscal            = _mm_and_ps(fscal,cutoff_mask);
1570
1571              /* Update vectorial force */
1572             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1573             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1574             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1575
1576             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1577             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1578             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1579
1580             }
1581
1582             /**************************
1583              * CALCULATE INTERACTIONS *
1584              **************************/
1585
1586             if (gmx_mm_any_lt(rsq23,rcutoff2))
1587             {
1588
1589             /* REACTION-FIELD ELECTROSTATICS */
1590             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
1591
1592             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
1593
1594             fscal            = felec;
1595
1596             fscal            = _mm_and_ps(fscal,cutoff_mask);
1597
1598              /* Update vectorial force */
1599             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1600             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1601             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1602
1603             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1604             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1605             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1606
1607             }
1608
1609             /**************************
1610              * CALCULATE INTERACTIONS *
1611              **************************/
1612
1613             if (gmx_mm_any_lt(rsq31,rcutoff2))
1614             {
1615
1616             /* REACTION-FIELD ELECTROSTATICS */
1617             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
1618
1619             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
1620
1621             fscal            = felec;
1622
1623             fscal            = _mm_and_ps(fscal,cutoff_mask);
1624
1625              /* Update vectorial force */
1626             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1627             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1628             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1629
1630             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1631             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1632             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1633
1634             }
1635
1636             /**************************
1637              * CALCULATE INTERACTIONS *
1638              **************************/
1639
1640             if (gmx_mm_any_lt(rsq32,rcutoff2))
1641             {
1642
1643             /* REACTION-FIELD ELECTROSTATICS */
1644             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
1645
1646             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
1647
1648             fscal            = felec;
1649
1650             fscal            = _mm_and_ps(fscal,cutoff_mask);
1651
1652              /* Update vectorial force */
1653             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1654             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1655             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1656
1657             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1658             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1659             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1660
1661             }
1662
1663             /**************************
1664              * CALCULATE INTERACTIONS *
1665              **************************/
1666
1667             if (gmx_mm_any_lt(rsq33,rcutoff2))
1668             {
1669
1670             /* REACTION-FIELD ELECTROSTATICS */
1671             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
1672
1673             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
1674
1675             fscal            = felec;
1676
1677             fscal            = _mm_and_ps(fscal,cutoff_mask);
1678
1679              /* Update vectorial force */
1680             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1681             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1682             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1683
1684             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1685             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1686             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1687
1688             }
1689
1690             fjptrA             = f+j_coord_offsetA;
1691             fjptrB             = f+j_coord_offsetB;
1692             fjptrC             = f+j_coord_offsetC;
1693             fjptrD             = f+j_coord_offsetD;
1694
1695             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1696                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1697                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1698
1699             /* Inner loop uses 333 flops */
1700         }
1701
1702         if(jidx<j_index_end)
1703         {
1704
1705             /* Get j neighbor index, and coordinate index */
1706             jnrlistA         = jjnr[jidx];
1707             jnrlistB         = jjnr[jidx+1];
1708             jnrlistC         = jjnr[jidx+2];
1709             jnrlistD         = jjnr[jidx+3];
1710             /* Sign of each element will be negative for non-real atoms.
1711              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1712              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1713              */
1714             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1715             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1716             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1717             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1718             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1719             j_coord_offsetA  = DIM*jnrA;
1720             j_coord_offsetB  = DIM*jnrB;
1721             j_coord_offsetC  = DIM*jnrC;
1722             j_coord_offsetD  = DIM*jnrD;
1723
1724             /* load j atom coordinates */
1725             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1726                                               x+j_coord_offsetC,x+j_coord_offsetD,
1727                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1728                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1729
1730             /* Calculate displacement vector */
1731             dx00             = _mm_sub_ps(ix0,jx0);
1732             dy00             = _mm_sub_ps(iy0,jy0);
1733             dz00             = _mm_sub_ps(iz0,jz0);
1734             dx11             = _mm_sub_ps(ix1,jx1);
1735             dy11             = _mm_sub_ps(iy1,jy1);
1736             dz11             = _mm_sub_ps(iz1,jz1);
1737             dx12             = _mm_sub_ps(ix1,jx2);
1738             dy12             = _mm_sub_ps(iy1,jy2);
1739             dz12             = _mm_sub_ps(iz1,jz2);
1740             dx13             = _mm_sub_ps(ix1,jx3);
1741             dy13             = _mm_sub_ps(iy1,jy3);
1742             dz13             = _mm_sub_ps(iz1,jz3);
1743             dx21             = _mm_sub_ps(ix2,jx1);
1744             dy21             = _mm_sub_ps(iy2,jy1);
1745             dz21             = _mm_sub_ps(iz2,jz1);
1746             dx22             = _mm_sub_ps(ix2,jx2);
1747             dy22             = _mm_sub_ps(iy2,jy2);
1748             dz22             = _mm_sub_ps(iz2,jz2);
1749             dx23             = _mm_sub_ps(ix2,jx3);
1750             dy23             = _mm_sub_ps(iy2,jy3);
1751             dz23             = _mm_sub_ps(iz2,jz3);
1752             dx31             = _mm_sub_ps(ix3,jx1);
1753             dy31             = _mm_sub_ps(iy3,jy1);
1754             dz31             = _mm_sub_ps(iz3,jz1);
1755             dx32             = _mm_sub_ps(ix3,jx2);
1756             dy32             = _mm_sub_ps(iy3,jy2);
1757             dz32             = _mm_sub_ps(iz3,jz2);
1758             dx33             = _mm_sub_ps(ix3,jx3);
1759             dy33             = _mm_sub_ps(iy3,jy3);
1760             dz33             = _mm_sub_ps(iz3,jz3);
1761
1762             /* Calculate squared distance and things based on it */
1763             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1764             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1765             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1766             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1767             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1768             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1769             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1770             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1771             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1772             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1773
1774             rinv11           = avx128fma_invsqrt_f(rsq11);
1775             rinv12           = avx128fma_invsqrt_f(rsq12);
1776             rinv13           = avx128fma_invsqrt_f(rsq13);
1777             rinv21           = avx128fma_invsqrt_f(rsq21);
1778             rinv22           = avx128fma_invsqrt_f(rsq22);
1779             rinv23           = avx128fma_invsqrt_f(rsq23);
1780             rinv31           = avx128fma_invsqrt_f(rsq31);
1781             rinv32           = avx128fma_invsqrt_f(rsq32);
1782             rinv33           = avx128fma_invsqrt_f(rsq33);
1783
1784             rinvsq00         = avx128fma_inv_f(rsq00);
1785             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1786             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1787             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1788             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1789             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1790             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1791             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1792             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1793             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1794
1795             fjx0             = _mm_setzero_ps();
1796             fjy0             = _mm_setzero_ps();
1797             fjz0             = _mm_setzero_ps();
1798             fjx1             = _mm_setzero_ps();
1799             fjy1             = _mm_setzero_ps();
1800             fjz1             = _mm_setzero_ps();
1801             fjx2             = _mm_setzero_ps();
1802             fjy2             = _mm_setzero_ps();
1803             fjz2             = _mm_setzero_ps();
1804             fjx3             = _mm_setzero_ps();
1805             fjy3             = _mm_setzero_ps();
1806             fjz3             = _mm_setzero_ps();
1807
1808             /**************************
1809              * CALCULATE INTERACTIONS *
1810              **************************/
1811
1812             if (gmx_mm_any_lt(rsq00,rcutoff2))
1813             {
1814
1815             /* LENNARD-JONES DISPERSION/REPULSION */
1816
1817             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1818             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1819
1820             cutoff_mask      = _mm_cmplt_ps(rsq00,rcutoff2);
1821
1822             fscal            = fvdw;
1823
1824             fscal            = _mm_and_ps(fscal,cutoff_mask);
1825
1826             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1827
1828              /* Update vectorial force */
1829             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1830             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1831             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1832
1833             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1834             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1835             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1836
1837             }
1838
1839             /**************************
1840              * CALCULATE INTERACTIONS *
1841              **************************/
1842
1843             if (gmx_mm_any_lt(rsq11,rcutoff2))
1844             {
1845
1846             /* REACTION-FIELD ELECTROSTATICS */
1847             felec            = _mm_mul_ps(qq11,_mm_msub_ps(rinv11,rinvsq11,krf2));
1848
1849             cutoff_mask      = _mm_cmplt_ps(rsq11,rcutoff2);
1850
1851             fscal            = felec;
1852
1853             fscal            = _mm_and_ps(fscal,cutoff_mask);
1854
1855             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1856
1857              /* Update vectorial force */
1858             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1859             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1860             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1861
1862             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1863             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1864             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1865
1866             }
1867
1868             /**************************
1869              * CALCULATE INTERACTIONS *
1870              **************************/
1871
1872             if (gmx_mm_any_lt(rsq12,rcutoff2))
1873             {
1874
1875             /* REACTION-FIELD ELECTROSTATICS */
1876             felec            = _mm_mul_ps(qq12,_mm_msub_ps(rinv12,rinvsq12,krf2));
1877
1878             cutoff_mask      = _mm_cmplt_ps(rsq12,rcutoff2);
1879
1880             fscal            = felec;
1881
1882             fscal            = _mm_and_ps(fscal,cutoff_mask);
1883
1884             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1885
1886              /* Update vectorial force */
1887             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1888             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1889             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1890
1891             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1892             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1893             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1894
1895             }
1896
1897             /**************************
1898              * CALCULATE INTERACTIONS *
1899              **************************/
1900
1901             if (gmx_mm_any_lt(rsq13,rcutoff2))
1902             {
1903
1904             /* REACTION-FIELD ELECTROSTATICS */
1905             felec            = _mm_mul_ps(qq13,_mm_msub_ps(rinv13,rinvsq13,krf2));
1906
1907             cutoff_mask      = _mm_cmplt_ps(rsq13,rcutoff2);
1908
1909             fscal            = felec;
1910
1911             fscal            = _mm_and_ps(fscal,cutoff_mask);
1912
1913             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1914
1915              /* Update vectorial force */
1916             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1917             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1918             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1919
1920             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1921             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1922             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1923
1924             }
1925
1926             /**************************
1927              * CALCULATE INTERACTIONS *
1928              **************************/
1929
1930             if (gmx_mm_any_lt(rsq21,rcutoff2))
1931             {
1932
1933             /* REACTION-FIELD ELECTROSTATICS */
1934             felec            = _mm_mul_ps(qq21,_mm_msub_ps(rinv21,rinvsq21,krf2));
1935
1936             cutoff_mask      = _mm_cmplt_ps(rsq21,rcutoff2);
1937
1938             fscal            = felec;
1939
1940             fscal            = _mm_and_ps(fscal,cutoff_mask);
1941
1942             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1943
1944              /* Update vectorial force */
1945             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1946             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1947             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1948
1949             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1950             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1951             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1952
1953             }
1954
1955             /**************************
1956              * CALCULATE INTERACTIONS *
1957              **************************/
1958
1959             if (gmx_mm_any_lt(rsq22,rcutoff2))
1960             {
1961
1962             /* REACTION-FIELD ELECTROSTATICS */
1963             felec            = _mm_mul_ps(qq22,_mm_msub_ps(rinv22,rinvsq22,krf2));
1964
1965             cutoff_mask      = _mm_cmplt_ps(rsq22,rcutoff2);
1966
1967             fscal            = felec;
1968
1969             fscal            = _mm_and_ps(fscal,cutoff_mask);
1970
1971             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1972
1973              /* Update vectorial force */
1974             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1975             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1976             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1977
1978             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1979             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1980             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1981
1982             }
1983
1984             /**************************
1985              * CALCULATE INTERACTIONS *
1986              **************************/
1987
1988             if (gmx_mm_any_lt(rsq23,rcutoff2))
1989             {
1990
1991             /* REACTION-FIELD ELECTROSTATICS */
1992             felec            = _mm_mul_ps(qq23,_mm_msub_ps(rinv23,rinvsq23,krf2));
1993
1994             cutoff_mask      = _mm_cmplt_ps(rsq23,rcutoff2);
1995
1996             fscal            = felec;
1997
1998             fscal            = _mm_and_ps(fscal,cutoff_mask);
1999
2000             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2001
2002              /* Update vectorial force */
2003             fix2             = _mm_macc_ps(dx23,fscal,fix2);
2004             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
2005             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
2006
2007             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
2008             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
2009             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
2010
2011             }
2012
2013             /**************************
2014              * CALCULATE INTERACTIONS *
2015              **************************/
2016
2017             if (gmx_mm_any_lt(rsq31,rcutoff2))
2018             {
2019
2020             /* REACTION-FIELD ELECTROSTATICS */
2021             felec            = _mm_mul_ps(qq31,_mm_msub_ps(rinv31,rinvsq31,krf2));
2022
2023             cutoff_mask      = _mm_cmplt_ps(rsq31,rcutoff2);
2024
2025             fscal            = felec;
2026
2027             fscal            = _mm_and_ps(fscal,cutoff_mask);
2028
2029             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2030
2031              /* Update vectorial force */
2032             fix3             = _mm_macc_ps(dx31,fscal,fix3);
2033             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
2034             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
2035
2036             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
2037             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
2038             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
2039
2040             }
2041
2042             /**************************
2043              * CALCULATE INTERACTIONS *
2044              **************************/
2045
2046             if (gmx_mm_any_lt(rsq32,rcutoff2))
2047             {
2048
2049             /* REACTION-FIELD ELECTROSTATICS */
2050             felec            = _mm_mul_ps(qq32,_mm_msub_ps(rinv32,rinvsq32,krf2));
2051
2052             cutoff_mask      = _mm_cmplt_ps(rsq32,rcutoff2);
2053
2054             fscal            = felec;
2055
2056             fscal            = _mm_and_ps(fscal,cutoff_mask);
2057
2058             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2059
2060              /* Update vectorial force */
2061             fix3             = _mm_macc_ps(dx32,fscal,fix3);
2062             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
2063             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
2064
2065             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
2066             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
2067             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
2068
2069             }
2070
2071             /**************************
2072              * CALCULATE INTERACTIONS *
2073              **************************/
2074
2075             if (gmx_mm_any_lt(rsq33,rcutoff2))
2076             {
2077
2078             /* REACTION-FIELD ELECTROSTATICS */
2079             felec            = _mm_mul_ps(qq33,_mm_msub_ps(rinv33,rinvsq33,krf2));
2080
2081             cutoff_mask      = _mm_cmplt_ps(rsq33,rcutoff2);
2082
2083             fscal            = felec;
2084
2085             fscal            = _mm_and_ps(fscal,cutoff_mask);
2086
2087             fscal            = _mm_andnot_ps(dummy_mask,fscal);
2088
2089              /* Update vectorial force */
2090             fix3             = _mm_macc_ps(dx33,fscal,fix3);
2091             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
2092             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
2093
2094             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
2095             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
2096             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
2097
2098             }
2099
2100             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
2101             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
2102             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
2103             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
2104
2105             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
2106                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
2107                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
2108
2109             /* Inner loop uses 333 flops */
2110         }
2111
2112         /* End of innermost loop */
2113
2114         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
2115                                               f+i_coord_offset,fshift+i_shift_offset);
2116
2117         /* Increment number of inner iterations */
2118         inneriter                  += j_index_end - j_index_start;
2119
2120         /* Outer loop uses 24 flops */
2121     }
2122
2123     /* Increment number of outer iterations */
2124     outeriter        += nri;
2125
2126     /* Update outer/inner flops */
2127
2128     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*333);
2129 }