src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecCoul_VdwLJ_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_ElecCoul_VdwLJ_GeomW4W4_VF_avx_128_fma_single
  51  * Electrostatics interaction: Coulomb
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Water4-Water4
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecCoul_VdwLJ_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     nvdwtype         = fr->ntype;
 134     vdwparam         = fr->nbfp;
 135     vdwtype          = mdatoms->typeA;
 136
 137     /* Setup water-specific parameters */
 138     inr              = nlist->iinr[0];
 139     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
 140     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
 141     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
 142     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 143
 144     jq1              = _mm_set1_ps(charge[inr+1]);
 145     jq2              = _mm_set1_ps(charge[inr+2]);
 146     jq3              = _mm_set1_ps(charge[inr+3]);
 147     vdwjidx0A        = 2*vdwtype[inr+0];
 148     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
 149     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
 150     qq11             = _mm_mul_ps(iq1,jq1);
 151     qq12             = _mm_mul_ps(iq1,jq2);
 152     qq13             = _mm_mul_ps(iq1,jq3);
 153     qq21             = _mm_mul_ps(iq2,jq1);
 154     qq22             = _mm_mul_ps(iq2,jq2);
 155     qq23             = _mm_mul_ps(iq2,jq3);
 156     qq31             = _mm_mul_ps(iq3,jq1);
 157     qq32             = _mm_mul_ps(iq3,jq2);
 158     qq33             = _mm_mul_ps(iq3,jq3);
 159
 160     /* Avoid stupid compiler warnings */
 161     jnrA = jnrB = jnrC = jnrD = 0;
 162     j_coord_offsetA = 0;
 163     j_coord_offsetB = 0;
 164     j_coord_offsetC = 0;
 165     j_coord_offsetD = 0;
 166
 167     outeriter        = 0;
 168     inneriter        = 0;
 169
 170     for(iidx=0;iidx<4*DIM;iidx++)
 171     {
 172         scratch[iidx] = 0.0;
 173     }
 174
 175     /* Start outer loop over neighborlists */
 176     for(iidx=0; iidx<nri; iidx++)
 177     {
 178         /* Load shift vector for this list */
 179         i_shift_offset   = DIM*shiftidx[iidx];
 180
 181         /* Load limits for loop over neighbors */
 182         j_index_start    = jindex[iidx];
 183         j_index_end      = jindex[iidx+1];
 184
 185         /* Get outer coordinate index */
 186         inr              = iinr[iidx];
 187         i_coord_offset   = DIM*inr;
 188
 189         /* Load i particle coords and add shift vector */
 190         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
 191                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 192
 193         fix0             = _mm_setzero_ps();
 194         fiy0             = _mm_setzero_ps();
 195         fiz0             = _mm_setzero_ps();
 196         fix1             = _mm_setzero_ps();
 197         fiy1             = _mm_setzero_ps();
 198         fiz1             = _mm_setzero_ps();
 199         fix2             = _mm_setzero_ps();
 200         fiy2             = _mm_setzero_ps();
 201         fiz2             = _mm_setzero_ps();
 202         fix3             = _mm_setzero_ps();
 203         fiy3             = _mm_setzero_ps();
 204         fiz3             = _mm_setzero_ps();
 205
 206         /* Reset potential sums */
 207         velecsum         = _mm_setzero_ps();
 208         vvdwsum          = _mm_setzero_ps();
 209
 210         /* Start inner kernel loop */
 211         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
 212         {
 213
 214             /* Get j neighbor index, and coordinate index */
 215             jnrA             = jjnr[jidx];
 216             jnrB             = jjnr[jidx+1];
 217             jnrC             = jjnr[jidx+2];
 218             jnrD             = jjnr[jidx+3];
 219             j_coord_offsetA  = DIM*jnrA;
 220             j_coord_offsetB  = DIM*jnrB;
 221             j_coord_offsetC  = DIM*jnrC;
 222             j_coord_offsetD  = DIM*jnrD;
 223
 224             /* load j atom coordinates */
 225             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 226                                               x+j_coord_offsetC,x+j_coord_offsetD,
 227                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 228                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 229
 230             /* Calculate displacement vector */
 231             dx00             = _mm_sub_ps(ix0,jx0);
 232             dy00             = _mm_sub_ps(iy0,jy0);
 233             dz00             = _mm_sub_ps(iz0,jz0);
 234             dx11             = _mm_sub_ps(ix1,jx1);
 235             dy11             = _mm_sub_ps(iy1,jy1);
 236             dz11             = _mm_sub_ps(iz1,jz1);
 237             dx12             = _mm_sub_ps(ix1,jx2);
 238             dy12             = _mm_sub_ps(iy1,jy2);
 239             dz12             = _mm_sub_ps(iz1,jz2);
 240             dx13             = _mm_sub_ps(ix1,jx3);
 241             dy13             = _mm_sub_ps(iy1,jy3);
 242             dz13             = _mm_sub_ps(iz1,jz3);
 243             dx21             = _mm_sub_ps(ix2,jx1);
 244             dy21             = _mm_sub_ps(iy2,jy1);
 245             dz21             = _mm_sub_ps(iz2,jz1);
 246             dx22             = _mm_sub_ps(ix2,jx2);
 247             dy22             = _mm_sub_ps(iy2,jy2);
 248             dz22             = _mm_sub_ps(iz2,jz2);
 249             dx23             = _mm_sub_ps(ix2,jx3);
 250             dy23             = _mm_sub_ps(iy2,jy3);
 251             dz23             = _mm_sub_ps(iz2,jz3);
 252             dx31             = _mm_sub_ps(ix3,jx1);
 253             dy31             = _mm_sub_ps(iy3,jy1);
 254             dz31             = _mm_sub_ps(iz3,jz1);
 255             dx32             = _mm_sub_ps(ix3,jx2);
 256             dy32             = _mm_sub_ps(iy3,jy2);
 257             dz32             = _mm_sub_ps(iz3,jz2);
 258             dx33             = _mm_sub_ps(ix3,jx3);
 259             dy33             = _mm_sub_ps(iy3,jy3);
 260             dz33             = _mm_sub_ps(iz3,jz3);
 261
 262             /* Calculate squared distance and things based on it */
 263             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 264             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 265             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 266             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 267             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 268             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 269             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 270             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 271             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 272             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 273
 274             rinv11           = avx128fma_invsqrt_f(rsq11);
 275             rinv12           = avx128fma_invsqrt_f(rsq12);
 276             rinv13           = avx128fma_invsqrt_f(rsq13);
 277             rinv21           = avx128fma_invsqrt_f(rsq21);
 278             rinv22           = avx128fma_invsqrt_f(rsq22);
 279             rinv23           = avx128fma_invsqrt_f(rsq23);
 280             rinv31           = avx128fma_invsqrt_f(rsq31);
 281             rinv32           = avx128fma_invsqrt_f(rsq32);
 282             rinv33           = avx128fma_invsqrt_f(rsq33);
 283
 284             rinvsq00         = avx128fma_inv_f(rsq00);
 285             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 286             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 287             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 288             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 289             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 290             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 291             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 292             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 293             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 294
 295             fjx0             = _mm_setzero_ps();
 296             fjy0             = _mm_setzero_ps();
 297             fjz0             = _mm_setzero_ps();
 298             fjx1             = _mm_setzero_ps();
 299             fjy1             = _mm_setzero_ps();
 300             fjz1             = _mm_setzero_ps();
 301             fjx2             = _mm_setzero_ps();
 302             fjy2             = _mm_setzero_ps();
 303             fjz2             = _mm_setzero_ps();
 304             fjx3             = _mm_setzero_ps();
 305             fjy3             = _mm_setzero_ps();
 306             fjz3             = _mm_setzero_ps();
 307
 308             /**************************
 309              * CALCULATE INTERACTIONS *
 310              **************************/
 311
 312             /* LENNARD-JONES DISPERSION/REPULSION */
 313
 314             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 315             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 316             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 317             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 318             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 319
 320             /* Update potential sum for this i atom from the interaction with this j atom. */
 321             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 322
 323             fscal            = fvdw;
 324
 325              /* Update vectorial force */
 326             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 327             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 328             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 329
 330             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 331             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 332             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 333
 334             /**************************
 335              * CALCULATE INTERACTIONS *
 336              **************************/
 337
 338             /* COULOMB ELECTROSTATICS */
 339             velec            = _mm_mul_ps(qq11,rinv11);
 340             felec            = _mm_mul_ps(velec,rinvsq11);
 341
 342             /* Update potential sum for this i atom from the interaction with this j atom. */
 343             velecsum         = _mm_add_ps(velecsum,velec);
 344
 345             fscal            = felec;
 346
 347              /* Update vectorial force */
 348             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 349             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 350             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 351
 352             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 353             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 354             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 355
 356             /**************************
 357              * CALCULATE INTERACTIONS *
 358              **************************/
 359
 360             /* COULOMB ELECTROSTATICS */
 361             velec            = _mm_mul_ps(qq12,rinv12);
 362             felec            = _mm_mul_ps(velec,rinvsq12);
 363
 364             /* Update potential sum for this i atom from the interaction with this j atom. */
 365             velecsum         = _mm_add_ps(velecsum,velec);
 366
 367             fscal            = felec;
 368
 369              /* Update vectorial force */
 370             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 371             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 372             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 373
 374             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 375             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 376             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 377
 378             /**************************
 379              * CALCULATE INTERACTIONS *
 380              **************************/
 381
 382             /* COULOMB ELECTROSTATICS */
 383             velec            = _mm_mul_ps(qq13,rinv13);
 384             felec            = _mm_mul_ps(velec,rinvsq13);
 385
 386             /* Update potential sum for this i atom from the interaction with this j atom. */
 387             velecsum         = _mm_add_ps(velecsum,velec);
 388
 389             fscal            = felec;
 390
 391              /* Update vectorial force */
 392             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 393             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 394             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 395
 396             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 397             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 398             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 399
 400             /**************************
 401              * CALCULATE INTERACTIONS *
 402              **************************/
 403
 404             /* COULOMB ELECTROSTATICS */
 405             velec            = _mm_mul_ps(qq21,rinv21);
 406             felec            = _mm_mul_ps(velec,rinvsq21);
 407
 408             /* Update potential sum for this i atom from the interaction with this j atom. */
 409             velecsum         = _mm_add_ps(velecsum,velec);
 410
 411             fscal            = felec;
 412
 413              /* Update vectorial force */
 414             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 415             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 416             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 417
 418             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 419             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 420             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 421
 422             /**************************
 423              * CALCULATE INTERACTIONS *
 424              **************************/
 425
 426             /* COULOMB ELECTROSTATICS */
 427             velec            = _mm_mul_ps(qq22,rinv22);
 428             felec            = _mm_mul_ps(velec,rinvsq22);
 429
 430             /* Update potential sum for this i atom from the interaction with this j atom. */
 431             velecsum         = _mm_add_ps(velecsum,velec);
 432
 433             fscal            = felec;
 434
 435              /* Update vectorial force */
 436             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 437             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 438             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 439
 440             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 441             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 442             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 443
 444             /**************************
 445              * CALCULATE INTERACTIONS *
 446              **************************/
 447
 448             /* COULOMB ELECTROSTATICS */
 449             velec            = _mm_mul_ps(qq23,rinv23);
 450             felec            = _mm_mul_ps(velec,rinvsq23);
 451
 452             /* Update potential sum for this i atom from the interaction with this j atom. */
 453             velecsum         = _mm_add_ps(velecsum,velec);
 454
 455             fscal            = felec;
 456
 457              /* Update vectorial force */
 458             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 459             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 460             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 461
 462             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 463             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 464             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 465
 466             /**************************
 467              * CALCULATE INTERACTIONS *
 468              **************************/
 469
 470             /* COULOMB ELECTROSTATICS */
 471             velec            = _mm_mul_ps(qq31,rinv31);
 472             felec            = _mm_mul_ps(velec,rinvsq31);
 473
 474             /* Update potential sum for this i atom from the interaction with this j atom. */
 475             velecsum         = _mm_add_ps(velecsum,velec);
 476
 477             fscal            = felec;
 478
 479              /* Update vectorial force */
 480             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 481             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 482             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 483
 484             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 485             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 486             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 487
 488             /**************************
 489              * CALCULATE INTERACTIONS *
 490              **************************/
 491
 492             /* COULOMB ELECTROSTATICS */
 493             velec            = _mm_mul_ps(qq32,rinv32);
 494             felec            = _mm_mul_ps(velec,rinvsq32);
 495
 496             /* Update potential sum for this i atom from the interaction with this j atom. */
 497             velecsum         = _mm_add_ps(velecsum,velec);
 498
 499             fscal            = felec;
 500
 501              /* Update vectorial force */
 502             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 503             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 504             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 505
 506             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 507             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 508             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 509
 510             /**************************
 511              * CALCULATE INTERACTIONS *
 512              **************************/
 513
 514             /* COULOMB ELECTROSTATICS */
 515             velec            = _mm_mul_ps(qq33,rinv33);
 516             felec            = _mm_mul_ps(velec,rinvsq33);
 517
 518             /* Update potential sum for this i atom from the interaction with this j atom. */
 519             velecsum         = _mm_add_ps(velecsum,velec);
 520
 521             fscal            = felec;
 522
 523              /* Update vectorial force */
 524             fix3             = _mm_macc_ps(dx33,fscal,fix3);
 525             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
 526             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
 527
 528             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
 529             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
 530             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
 531
 532             fjptrA             = f+j_coord_offsetA;
 533             fjptrB             = f+j_coord_offsetB;
 534             fjptrC             = f+j_coord_offsetC;
 535             fjptrD             = f+j_coord_offsetD;
 536
 537             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 538                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 539                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 540
 541             /* Inner loop uses 317 flops */
 542         }
 543
 544         if(jidx<j_index_end)
 545         {
 546
 547             /* Get j neighbor index, and coordinate index */
 548             jnrlistA         = jjnr[jidx];
 549             jnrlistB         = jjnr[jidx+1];
 550             jnrlistC         = jjnr[jidx+2];
 551             jnrlistD         = jjnr[jidx+3];
 552             /* Sign of each element will be negative for non-real atoms.
 553              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
 554              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
 555              */
 556             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 557             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
 558             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
 559             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
 560             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
 561             j_coord_offsetA  = DIM*jnrA;
 562             j_coord_offsetB  = DIM*jnrB;
 563             j_coord_offsetC  = DIM*jnrC;
 564             j_coord_offsetD  = DIM*jnrD;
 565
 566             /* load j atom coordinates */
 567             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
 568                                               x+j_coord_offsetC,x+j_coord_offsetD,
 569                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 570                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 571
 572             /* Calculate displacement vector */
 573             dx00             = _mm_sub_ps(ix0,jx0);
 574             dy00             = _mm_sub_ps(iy0,jy0);
 575             dz00             = _mm_sub_ps(iz0,jz0);
 576             dx11             = _mm_sub_ps(ix1,jx1);
 577             dy11             = _mm_sub_ps(iy1,jy1);
 578             dz11             = _mm_sub_ps(iz1,jz1);
 579             dx12             = _mm_sub_ps(ix1,jx2);
 580             dy12             = _mm_sub_ps(iy1,jy2);
 581             dz12             = _mm_sub_ps(iz1,jz2);
 582             dx13             = _mm_sub_ps(ix1,jx3);
 583             dy13             = _mm_sub_ps(iy1,jy3);
 584             dz13             = _mm_sub_ps(iz1,jz3);
 585             dx21             = _mm_sub_ps(ix2,jx1);
 586             dy21             = _mm_sub_ps(iy2,jy1);
 587             dz21             = _mm_sub_ps(iz2,jz1);
 588             dx22             = _mm_sub_ps(ix2,jx2);
 589             dy22             = _mm_sub_ps(iy2,jy2);
 590             dz22             = _mm_sub_ps(iz2,jz2);
 591             dx23             = _mm_sub_ps(ix2,jx3);
 592             dy23             = _mm_sub_ps(iy2,jy3);
 593             dz23             = _mm_sub_ps(iz2,jz3);
 594             dx31             = _mm_sub_ps(ix3,jx1);
 595             dy31             = _mm_sub_ps(iy3,jy1);
 596             dz31             = _mm_sub_ps(iz3,jz1);
 597             dx32             = _mm_sub_ps(ix3,jx2);
 598             dy32             = _mm_sub_ps(iy3,jy2);
 599             dz32             = _mm_sub_ps(iz3,jz2);
 600             dx33             = _mm_sub_ps(ix3,jx3);
 601             dy33             = _mm_sub_ps(iy3,jy3);
 602             dz33             = _mm_sub_ps(iz3,jz3);
 603
 604             /* Calculate squared distance and things based on it */
 605             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
 606             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
 607             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
 608             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
 609             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
 610             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
 611             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
 612             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
 613             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
 614             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
 615
 616             rinv11           = avx128fma_invsqrt_f(rsq11);
 617             rinv12           = avx128fma_invsqrt_f(rsq12);
 618             rinv13           = avx128fma_invsqrt_f(rsq13);
 619             rinv21           = avx128fma_invsqrt_f(rsq21);
 620             rinv22           = avx128fma_invsqrt_f(rsq22);
 621             rinv23           = avx128fma_invsqrt_f(rsq23);
 622             rinv31           = avx128fma_invsqrt_f(rsq31);
 623             rinv32           = avx128fma_invsqrt_f(rsq32);
 624             rinv33           = avx128fma_invsqrt_f(rsq33);
 625
 626             rinvsq00         = avx128fma_inv_f(rsq00);
 627             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
 628             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
 629             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
 630             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
 631             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
 632             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
 633             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
 634             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
 635             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
 636
 637             fjx0             = _mm_setzero_ps();
 638             fjy0             = _mm_setzero_ps();
 639             fjz0             = _mm_setzero_ps();
 640             fjx1             = _mm_setzero_ps();
 641             fjy1             = _mm_setzero_ps();
 642             fjz1             = _mm_setzero_ps();
 643             fjx2             = _mm_setzero_ps();
 644             fjy2             = _mm_setzero_ps();
 645             fjz2             = _mm_setzero_ps();
 646             fjx3             = _mm_setzero_ps();
 647             fjy3             = _mm_setzero_ps();
 648             fjz3             = _mm_setzero_ps();
 649
 650             /**************************
 651              * CALCULATE INTERACTIONS *
 652              **************************/
 653
 654             /* LENNARD-JONES DISPERSION/REPULSION */
 655
 656             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
 657             vvdw6            = _mm_mul_ps(c6_00,rinvsix);
 658             vvdw12           = _mm_mul_ps(c12_00,_mm_mul_ps(rinvsix,rinvsix));
 659             vvdw             = _mm_msub_ps(vvdw12,one_twelfth,_mm_mul_ps(vvdw6,one_sixth));
 660             fvdw             = _mm_mul_ps(_mm_sub_ps(vvdw12,vvdw6),rinvsq00);
 661
 662             /* Update potential sum for this i atom from the interaction with this j atom. */
 663             vvdw             = _mm_andnot_ps(dummy_mask,vvdw);
 664             vvdwsum          = _mm_add_ps(vvdwsum,vvdw);
 665
 666             fscal            = fvdw;
 667
 668             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 669
 670              /* Update vectorial force */
 671             fix0             = _mm_macc_ps(dx00,fscal,fix0);
 672             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
 673             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
 674
 675             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
 676             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
 677             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
 678
 679             /**************************
 680              * CALCULATE INTERACTIONS *
 681              **************************/
 682
 683             /* COULOMB ELECTROSTATICS */
 684             velec            = _mm_mul_ps(qq11,rinv11);
 685             felec            = _mm_mul_ps(velec,rinvsq11);
 686
 687             /* Update potential sum for this i atom from the interaction with this j atom. */
 688             velec            = _mm_andnot_ps(dummy_mask,velec);
 689             velecsum         = _mm_add_ps(velecsum,velec);
 690
 691             fscal            = felec;
 692
 693             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 694
 695              /* Update vectorial force */
 696             fix1             = _mm_macc_ps(dx11,fscal,fix1);
 697             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
 698             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
 699
 700             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
 701             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
 702             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
 703
 704             /**************************
 705              * CALCULATE INTERACTIONS *
 706              **************************/
 707
 708             /* COULOMB ELECTROSTATICS */
 709             velec            = _mm_mul_ps(qq12,rinv12);
 710             felec            = _mm_mul_ps(velec,rinvsq12);
 711
 712             /* Update potential sum for this i atom from the interaction with this j atom. */
 713             velec            = _mm_andnot_ps(dummy_mask,velec);
 714             velecsum         = _mm_add_ps(velecsum,velec);
 715
 716             fscal            = felec;
 717
 718             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 719
 720              /* Update vectorial force */
 721             fix1             = _mm_macc_ps(dx12,fscal,fix1);
 722             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
 723             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
 724
 725             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
 726             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
 727             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
 728
 729             /**************************
 730              * CALCULATE INTERACTIONS *
 731              **************************/
 732
 733             /* COULOMB ELECTROSTATICS */
 734             velec            = _mm_mul_ps(qq13,rinv13);
 735             felec            = _mm_mul_ps(velec,rinvsq13);
 736
 737             /* Update potential sum for this i atom from the interaction with this j atom. */
 738             velec            = _mm_andnot_ps(dummy_mask,velec);
 739             velecsum         = _mm_add_ps(velecsum,velec);
 740
 741             fscal            = felec;
 742
 743             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 744
 745              /* Update vectorial force */
 746             fix1             = _mm_macc_ps(dx13,fscal,fix1);
 747             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
 748             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
 749
 750             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
 751             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
 752             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
 753
 754             /**************************
 755              * CALCULATE INTERACTIONS *
 756              **************************/
 757
 758             /* COULOMB ELECTROSTATICS */
 759             velec            = _mm_mul_ps(qq21,rinv21);
 760             felec            = _mm_mul_ps(velec,rinvsq21);
 761
 762             /* Update potential sum for this i atom from the interaction with this j atom. */
 763             velec            = _mm_andnot_ps(dummy_mask,velec);
 764             velecsum         = _mm_add_ps(velecsum,velec);
 765
 766             fscal            = felec;
 767
 768             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 769
 770              /* Update vectorial force */
 771             fix2             = _mm_macc_ps(dx21,fscal,fix2);
 772             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
 773             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
 774
 775             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
 776             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
 777             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
 778
 779             /**************************
 780              * CALCULATE INTERACTIONS *
 781              **************************/
 782
 783             /* COULOMB ELECTROSTATICS */
 784             velec            = _mm_mul_ps(qq22,rinv22);
 785             felec            = _mm_mul_ps(velec,rinvsq22);
 786
 787             /* Update potential sum for this i atom from the interaction with this j atom. */
 788             velec            = _mm_andnot_ps(dummy_mask,velec);
 789             velecsum         = _mm_add_ps(velecsum,velec);
 790
 791             fscal            = felec;
 792
 793             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 794
 795              /* Update vectorial force */
 796             fix2             = _mm_macc_ps(dx22,fscal,fix2);
 797             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
 798             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
 799
 800             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
 801             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
 802             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
 803
 804             /**************************
 805              * CALCULATE INTERACTIONS *
 806              **************************/
 807
 808             /* COULOMB ELECTROSTATICS */
 809             velec            = _mm_mul_ps(qq23,rinv23);
 810             felec            = _mm_mul_ps(velec,rinvsq23);
 811
 812             /* Update potential sum for this i atom from the interaction with this j atom. */
 813             velec            = _mm_andnot_ps(dummy_mask,velec);
 814             velecsum         = _mm_add_ps(velecsum,velec);
 815
 816             fscal            = felec;
 817
 818             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 819
 820              /* Update vectorial force */
 821             fix2             = _mm_macc_ps(dx23,fscal,fix2);
 822             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
 823             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
 824
 825             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
 826             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
 827             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
 828
 829             /**************************
 830              * CALCULATE INTERACTIONS *
 831              **************************/
 832
 833             /* COULOMB ELECTROSTATICS */
 834             velec            = _mm_mul_ps(qq31,rinv31);
 835             felec            = _mm_mul_ps(velec,rinvsq31);
 836
 837             /* Update potential sum for this i atom from the interaction with this j atom. */
 838             velec            = _mm_andnot_ps(dummy_mask,velec);
 839             velecsum         = _mm_add_ps(velecsum,velec);
 840
 841             fscal            = felec;
 842
 843             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 844
 845              /* Update vectorial force */
 846             fix3             = _mm_macc_ps(dx31,fscal,fix3);
 847             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
 848             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
 849
 850             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
 851             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
 852             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
 853
 854             /**************************
 855              * CALCULATE INTERACTIONS *
 856              **************************/
 857
 858             /* COULOMB ELECTROSTATICS */
 859             velec            = _mm_mul_ps(qq32,rinv32);
 860             felec            = _mm_mul_ps(velec,rinvsq32);
 861
 862             /* Update potential sum for this i atom from the interaction with this j atom. */
 863             velec            = _mm_andnot_ps(dummy_mask,velec);
 864             velecsum         = _mm_add_ps(velecsum,velec);
 865
 866             fscal            = felec;
 867
 868             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 869
 870              /* Update vectorial force */
 871             fix3             = _mm_macc_ps(dx32,fscal,fix3);
 872             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
 873             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
 874
 875             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
 876             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
 877             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
 878
 879             /**************************
 880              * CALCULATE INTERACTIONS *
 881              **************************/
 882
 883             /* COULOMB ELECTROSTATICS */
 884             velec            = _mm_mul_ps(qq33,rinv33);
 885             felec            = _mm_mul_ps(velec,rinvsq33);
 886
 887             /* Update potential sum for this i atom from the interaction with this j atom. */
 888             velec            = _mm_andnot_ps(dummy_mask,velec);
 889             velecsum         = _mm_add_ps(velecsum,velec);
 890
 891             fscal            = felec;
 892
 893             fscal            = _mm_andnot_ps(dummy_mask,fscal);
 894
 895              /* Update vectorial force */
 896             fix3             = _mm_macc_ps(dx33,fscal,fix3);
 897             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
 898             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
 899
 900             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
 901             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
 902             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
 903
 904             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
 905             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
 906             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
 907             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
 908
 909             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
 910                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
 911                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 912
 913             /* Inner loop uses 317 flops */
 914         }
 915
 916         /* End of innermost loop */
 917
 918         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 919                                               f+i_coord_offset,fshift+i_shift_offset);
 920
 921         ggid                        = gid[iidx];
 922         /* Update potential energies */
 923         gmx_mm_update_1pot_ps(velecsum,kernel_data->energygrp_elec+ggid);
 924         gmx_mm_update_1pot_ps(vvdwsum,kernel_data->energygrp_vdw+ggid);
 925
 926         /* Increment number of inner iterations */
 927         inneriter                  += j_index_end - j_index_start;
 928
 929         /* Outer loop uses 26 flops */
 930     }
 931
 932     /* Increment number of outer iterations */
 933     outeriter        += nri;
 934
 935     /* Update outer/inner flops */
 936
 937     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*317);
 938 }
 939 /*
 940  * Gromacs nonbonded kernel:   nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_avx_128_fma_single
 941  * Electrostatics interaction: Coulomb
 942  * VdW interaction:            LennardJones
 943  * Geometry:                   Water4-Water4
 944  * Calculate force/pot:        Force
 945  */
 946 void
 947 nb_kernel_ElecCoul_VdwLJ_GeomW4W4_F_avx_128_fma_single
 948                     (t_nblist                    * gmx_restrict       nlist,
 949                      rvec                        * gmx_restrict          xx,
 950                      rvec                        * gmx_restrict          ff,
 951                      struct t_forcerec           * gmx_restrict          fr,
 952                      t_mdatoms                   * gmx_restrict     mdatoms,
 953                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 954                      t_nrnb                      * gmx_restrict        nrnb)
 955 {
 956     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 957      * just 0 for non-waters.
 958      * Suffixes A,B,C,D refer to j loop unrolling done with AVX_128, e.g. for the four different
 959      * jnr indices corresponding to data put in the four positions in the SIMD register.
 960      */
 961     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 962     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 963     int              jnrA,jnrB,jnrC,jnrD;
 964     int              jnrlistA,jnrlistB,jnrlistC,jnrlistD;
 965     int              j_coord_offsetA,j_coord_offsetB,j_coord_offsetC,j_coord_offsetD;
 966     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 967     real             rcutoff_scalar;
 968     real             *shiftvec,*fshift,*x,*f;
 969     real             *fjptrA,*fjptrB,*fjptrC,*fjptrD;
 970     real             scratch[4*DIM];
 971     __m128           fscal,rcutoff,rcutoff2,jidxall;
 972     int              vdwioffset0;
 973     __m128           ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 974     int              vdwioffset1;
 975     __m128           ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 976     int              vdwioffset2;
 977     __m128           ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 978     int              vdwioffset3;
 979     __m128           ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 980     int              vdwjidx0A,vdwjidx0B,vdwjidx0C,vdwjidx0D;
 981     __m128           jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 982     int              vdwjidx1A,vdwjidx1B,vdwjidx1C,vdwjidx1D;
 983     __m128           jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 984     int              vdwjidx2A,vdwjidx2B,vdwjidx2C,vdwjidx2D;
 985     __m128           jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 986     int              vdwjidx3A,vdwjidx3B,vdwjidx3C,vdwjidx3D;
 987     __m128           jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 988     __m128           dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 989     __m128           dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 990     __m128           dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 991     __m128           dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 992     __m128           dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 993     __m128           dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 994     __m128           dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 995     __m128           dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 996     __m128           dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 997     __m128           dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 998     __m128           velec,felec,velecsum,facel,crf,krf,krf2;
 999     real             *charge;
1000     int              nvdwtype;
1001     __m128           rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1002     int              *vdwtype;
1003     real             *vdwparam;
1004     __m128           one_sixth   = _mm_set1_ps(1.0/6.0);
1005     __m128           one_twelfth = _mm_set1_ps(1.0/12.0);
1006     __m128           dummy_mask,cutoff_mask;
1007     __m128           signbit = _mm_castsi128_ps( _mm_set1_epi32(0x80000000) );
1008     __m128           one     = _mm_set1_ps(1.0);
1009     __m128           two     = _mm_set1_ps(2.0);
1010     x                = xx[0];
1011     f                = ff[0];
1012
1013     nri              = nlist->nri;
1014     iinr             = nlist->iinr;
1015     jindex           = nlist->jindex;
1016     jjnr             = nlist->jjnr;
1017     shiftidx         = nlist->shift;
1018     gid              = nlist->gid;
1019     shiftvec         = fr->shift_vec[0];
1020     fshift           = fr->fshift[0];
1021     facel            = _mm_set1_ps(fr->ic->epsfac);
1022     charge           = mdatoms->chargeA;
1023     nvdwtype         = fr->ntype;
1024     vdwparam         = fr->nbfp;
1025     vdwtype          = mdatoms->typeA;
1026
1027     /* Setup water-specific parameters */
1028     inr              = nlist->iinr[0];
1029     iq1              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+1]));
1030     iq2              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+2]));
1031     iq3              = _mm_mul_ps(facel,_mm_set1_ps(charge[inr+3]));
1032     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1033
1034     jq1              = _mm_set1_ps(charge[inr+1]);
1035     jq2              = _mm_set1_ps(charge[inr+2]);
1036     jq3              = _mm_set1_ps(charge[inr+3]);
1037     vdwjidx0A        = 2*vdwtype[inr+0];
1038     c6_00            = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A]);
1039     c12_00           = _mm_set1_ps(vdwparam[vdwioffset0+vdwjidx0A+1]);
1040     qq11             = _mm_mul_ps(iq1,jq1);
1041     qq12             = _mm_mul_ps(iq1,jq2);
1042     qq13             = _mm_mul_ps(iq1,jq3);
1043     qq21             = _mm_mul_ps(iq2,jq1);
1044     qq22             = _mm_mul_ps(iq2,jq2);
1045     qq23             = _mm_mul_ps(iq2,jq3);
1046     qq31             = _mm_mul_ps(iq3,jq1);
1047     qq32             = _mm_mul_ps(iq3,jq2);
1048     qq33             = _mm_mul_ps(iq3,jq3);
1049
1050     /* Avoid stupid compiler warnings */
1051     jnrA = jnrB = jnrC = jnrD = 0;
1052     j_coord_offsetA = 0;
1053     j_coord_offsetB = 0;
1054     j_coord_offsetC = 0;
1055     j_coord_offsetD = 0;
1056
1057     outeriter        = 0;
1058     inneriter        = 0;
1059
1060     for(iidx=0;iidx<4*DIM;iidx++)
1061     {
1062         scratch[iidx] = 0.0;
1063     }
1064
1065     /* Start outer loop over neighborlists */
1066     for(iidx=0; iidx<nri; iidx++)
1067     {
1068         /* Load shift vector for this list */
1069         i_shift_offset   = DIM*shiftidx[iidx];
1070
1071         /* Load limits for loop over neighbors */
1072         j_index_start    = jindex[iidx];
1073         j_index_end      = jindex[iidx+1];
1074
1075         /* Get outer coordinate index */
1076         inr              = iinr[iidx];
1077         i_coord_offset   = DIM*inr;
1078
1079         /* Load i particle coords and add shift vector */
1080         gmx_mm_load_shift_and_4rvec_broadcast_ps(shiftvec+i_shift_offset,x+i_coord_offset,
1081                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1082
1083         fix0             = _mm_setzero_ps();
1084         fiy0             = _mm_setzero_ps();
1085         fiz0             = _mm_setzero_ps();
1086         fix1             = _mm_setzero_ps();
1087         fiy1             = _mm_setzero_ps();
1088         fiz1             = _mm_setzero_ps();
1089         fix2             = _mm_setzero_ps();
1090         fiy2             = _mm_setzero_ps();
1091         fiz2             = _mm_setzero_ps();
1092         fix3             = _mm_setzero_ps();
1093         fiy3             = _mm_setzero_ps();
1094         fiz3             = _mm_setzero_ps();
1095
1096         /* Start inner kernel loop */
1097         for(jidx=j_index_start; jidx<j_index_end && jjnr[jidx+3]>=0; jidx+=4)
1098         {
1099
1100             /* Get j neighbor index, and coordinate index */
1101             jnrA             = jjnr[jidx];
1102             jnrB             = jjnr[jidx+1];
1103             jnrC             = jjnr[jidx+2];
1104             jnrD             = jjnr[jidx+3];
1105             j_coord_offsetA  = DIM*jnrA;
1106             j_coord_offsetB  = DIM*jnrB;
1107             j_coord_offsetC  = DIM*jnrC;
1108             j_coord_offsetD  = DIM*jnrD;
1109
1110             /* load j atom coordinates */
1111             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1112                                               x+j_coord_offsetC,x+j_coord_offsetD,
1113                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1114                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1115
1116             /* Calculate displacement vector */
1117             dx00             = _mm_sub_ps(ix0,jx0);
1118             dy00             = _mm_sub_ps(iy0,jy0);
1119             dz00             = _mm_sub_ps(iz0,jz0);
1120             dx11             = _mm_sub_ps(ix1,jx1);
1121             dy11             = _mm_sub_ps(iy1,jy1);
1122             dz11             = _mm_sub_ps(iz1,jz1);
1123             dx12             = _mm_sub_ps(ix1,jx2);
1124             dy12             = _mm_sub_ps(iy1,jy2);
1125             dz12             = _mm_sub_ps(iz1,jz2);
1126             dx13             = _mm_sub_ps(ix1,jx3);
1127             dy13             = _mm_sub_ps(iy1,jy3);
1128             dz13             = _mm_sub_ps(iz1,jz3);
1129             dx21             = _mm_sub_ps(ix2,jx1);
1130             dy21             = _mm_sub_ps(iy2,jy1);
1131             dz21             = _mm_sub_ps(iz2,jz1);
1132             dx22             = _mm_sub_ps(ix2,jx2);
1133             dy22             = _mm_sub_ps(iy2,jy2);
1134             dz22             = _mm_sub_ps(iz2,jz2);
1135             dx23             = _mm_sub_ps(ix2,jx3);
1136             dy23             = _mm_sub_ps(iy2,jy3);
1137             dz23             = _mm_sub_ps(iz2,jz3);
1138             dx31             = _mm_sub_ps(ix3,jx1);
1139             dy31             = _mm_sub_ps(iy3,jy1);
1140             dz31             = _mm_sub_ps(iz3,jz1);
1141             dx32             = _mm_sub_ps(ix3,jx2);
1142             dy32             = _mm_sub_ps(iy3,jy2);
1143             dz32             = _mm_sub_ps(iz3,jz2);
1144             dx33             = _mm_sub_ps(ix3,jx3);
1145             dy33             = _mm_sub_ps(iy3,jy3);
1146             dz33             = _mm_sub_ps(iz3,jz3);
1147
1148             /* Calculate squared distance and things based on it */
1149             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1150             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1151             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1152             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1153             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1154             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1155             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1156             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1157             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1158             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1159
1160             rinv11           = avx128fma_invsqrt_f(rsq11);
1161             rinv12           = avx128fma_invsqrt_f(rsq12);
1162             rinv13           = avx128fma_invsqrt_f(rsq13);
1163             rinv21           = avx128fma_invsqrt_f(rsq21);
1164             rinv22           = avx128fma_invsqrt_f(rsq22);
1165             rinv23           = avx128fma_invsqrt_f(rsq23);
1166             rinv31           = avx128fma_invsqrt_f(rsq31);
1167             rinv32           = avx128fma_invsqrt_f(rsq32);
1168             rinv33           = avx128fma_invsqrt_f(rsq33);
1169
1170             rinvsq00         = avx128fma_inv_f(rsq00);
1171             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1172             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1173             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1174             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1175             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1176             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1177             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1178             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1179             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1180
1181             fjx0             = _mm_setzero_ps();
1182             fjy0             = _mm_setzero_ps();
1183             fjz0             = _mm_setzero_ps();
1184             fjx1             = _mm_setzero_ps();
1185             fjy1             = _mm_setzero_ps();
1186             fjz1             = _mm_setzero_ps();
1187             fjx2             = _mm_setzero_ps();
1188             fjy2             = _mm_setzero_ps();
1189             fjz2             = _mm_setzero_ps();
1190             fjx3             = _mm_setzero_ps();
1191             fjy3             = _mm_setzero_ps();
1192             fjz3             = _mm_setzero_ps();
1193
1194             /**************************
1195              * CALCULATE INTERACTIONS *
1196              **************************/
1197
1198             /* LENNARD-JONES DISPERSION/REPULSION */
1199
1200             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1201             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1202
1203             fscal            = fvdw;
1204
1205              /* Update vectorial force */
1206             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1207             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1208             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1209
1210             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1211             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1212             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1213
1214             /**************************
1215              * CALCULATE INTERACTIONS *
1216              **************************/
1217
1218             /* COULOMB ELECTROSTATICS */
1219             velec            = _mm_mul_ps(qq11,rinv11);
1220             felec            = _mm_mul_ps(velec,rinvsq11);
1221
1222             fscal            = felec;
1223
1224              /* Update vectorial force */
1225             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1226             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1227             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1228
1229             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1230             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1231             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1232
1233             /**************************
1234              * CALCULATE INTERACTIONS *
1235              **************************/
1236
1237             /* COULOMB ELECTROSTATICS */
1238             velec            = _mm_mul_ps(qq12,rinv12);
1239             felec            = _mm_mul_ps(velec,rinvsq12);
1240
1241             fscal            = felec;
1242
1243              /* Update vectorial force */
1244             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1245             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1246             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1247
1248             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1249             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1250             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1251
1252             /**************************
1253              * CALCULATE INTERACTIONS *
1254              **************************/
1255
1256             /* COULOMB ELECTROSTATICS */
1257             velec            = _mm_mul_ps(qq13,rinv13);
1258             felec            = _mm_mul_ps(velec,rinvsq13);
1259
1260             fscal            = felec;
1261
1262              /* Update vectorial force */
1263             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1264             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1265             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1266
1267             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1268             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1269             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1270
1271             /**************************
1272              * CALCULATE INTERACTIONS *
1273              **************************/
1274
1275             /* COULOMB ELECTROSTATICS */
1276             velec            = _mm_mul_ps(qq21,rinv21);
1277             felec            = _mm_mul_ps(velec,rinvsq21);
1278
1279             fscal            = felec;
1280
1281              /* Update vectorial force */
1282             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1283             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1284             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1285
1286             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1287             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1288             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1289
1290             /**************************
1291              * CALCULATE INTERACTIONS *
1292              **************************/
1293
1294             /* COULOMB ELECTROSTATICS */
1295             velec            = _mm_mul_ps(qq22,rinv22);
1296             felec            = _mm_mul_ps(velec,rinvsq22);
1297
1298             fscal            = felec;
1299
1300              /* Update vectorial force */
1301             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1302             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1303             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1304
1305             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1306             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1307             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1308
1309             /**************************
1310              * CALCULATE INTERACTIONS *
1311              **************************/
1312
1313             /* COULOMB ELECTROSTATICS */
1314             velec            = _mm_mul_ps(qq23,rinv23);
1315             felec            = _mm_mul_ps(velec,rinvsq23);
1316
1317             fscal            = felec;
1318
1319              /* Update vectorial force */
1320             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1321             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1322             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1323
1324             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1325             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1326             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1327
1328             /**************************
1329              * CALCULATE INTERACTIONS *
1330              **************************/
1331
1332             /* COULOMB ELECTROSTATICS */
1333             velec            = _mm_mul_ps(qq31,rinv31);
1334             felec            = _mm_mul_ps(velec,rinvsq31);
1335
1336             fscal            = felec;
1337
1338              /* Update vectorial force */
1339             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1340             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1341             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1342
1343             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1344             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1345             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1346
1347             /**************************
1348              * CALCULATE INTERACTIONS *
1349              **************************/
1350
1351             /* COULOMB ELECTROSTATICS */
1352             velec            = _mm_mul_ps(qq32,rinv32);
1353             felec            = _mm_mul_ps(velec,rinvsq32);
1354
1355             fscal            = felec;
1356
1357              /* Update vectorial force */
1358             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1359             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1360             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1361
1362             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1363             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1364             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1365
1366             /**************************
1367              * CALCULATE INTERACTIONS *
1368              **************************/
1369
1370             /* COULOMB ELECTROSTATICS */
1371             velec            = _mm_mul_ps(qq33,rinv33);
1372             felec            = _mm_mul_ps(velec,rinvsq33);
1373
1374             fscal            = felec;
1375
1376              /* Update vectorial force */
1377             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1378             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1379             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1380
1381             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1382             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1383             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1384
1385             fjptrA             = f+j_coord_offsetA;
1386             fjptrB             = f+j_coord_offsetB;
1387             fjptrC             = f+j_coord_offsetC;
1388             fjptrD             = f+j_coord_offsetD;
1389
1390             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1391                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1392                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1393
1394             /* Inner loop uses 303 flops */
1395         }
1396
1397         if(jidx<j_index_end)
1398         {
1399
1400             /* Get j neighbor index, and coordinate index */
1401             jnrlistA         = jjnr[jidx];
1402             jnrlistB         = jjnr[jidx+1];
1403             jnrlistC         = jjnr[jidx+2];
1404             jnrlistD         = jjnr[jidx+3];
1405             /* Sign of each element will be negative for non-real atoms.
1406              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
1407              * so use it as val = _mm_andnot_ps(mask,val) to clear dummy entries.
1408              */
1409             dummy_mask = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
1410             jnrA       = (jnrlistA>=0) ? jnrlistA : 0;
1411             jnrB       = (jnrlistB>=0) ? jnrlistB : 0;
1412             jnrC       = (jnrlistC>=0) ? jnrlistC : 0;
1413             jnrD       = (jnrlistD>=0) ? jnrlistD : 0;
1414             j_coord_offsetA  = DIM*jnrA;
1415             j_coord_offsetB  = DIM*jnrB;
1416             j_coord_offsetC  = DIM*jnrC;
1417             j_coord_offsetD  = DIM*jnrD;
1418
1419             /* load j atom coordinates */
1420             gmx_mm_load_4rvec_4ptr_swizzle_ps(x+j_coord_offsetA,x+j_coord_offsetB,
1421                                               x+j_coord_offsetC,x+j_coord_offsetD,
1422                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1423                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1424
1425             /* Calculate displacement vector */
1426             dx00             = _mm_sub_ps(ix0,jx0);
1427             dy00             = _mm_sub_ps(iy0,jy0);
1428             dz00             = _mm_sub_ps(iz0,jz0);
1429             dx11             = _mm_sub_ps(ix1,jx1);
1430             dy11             = _mm_sub_ps(iy1,jy1);
1431             dz11             = _mm_sub_ps(iz1,jz1);
1432             dx12             = _mm_sub_ps(ix1,jx2);
1433             dy12             = _mm_sub_ps(iy1,jy2);
1434             dz12             = _mm_sub_ps(iz1,jz2);
1435             dx13             = _mm_sub_ps(ix1,jx3);
1436             dy13             = _mm_sub_ps(iy1,jy3);
1437             dz13             = _mm_sub_ps(iz1,jz3);
1438             dx21             = _mm_sub_ps(ix2,jx1);
1439             dy21             = _mm_sub_ps(iy2,jy1);
1440             dz21             = _mm_sub_ps(iz2,jz1);
1441             dx22             = _mm_sub_ps(ix2,jx2);
1442             dy22             = _mm_sub_ps(iy2,jy2);
1443             dz22             = _mm_sub_ps(iz2,jz2);
1444             dx23             = _mm_sub_ps(ix2,jx3);
1445             dy23             = _mm_sub_ps(iy2,jy3);
1446             dz23             = _mm_sub_ps(iz2,jz3);
1447             dx31             = _mm_sub_ps(ix3,jx1);
1448             dy31             = _mm_sub_ps(iy3,jy1);
1449             dz31             = _mm_sub_ps(iz3,jz1);
1450             dx32             = _mm_sub_ps(ix3,jx2);
1451             dy32             = _mm_sub_ps(iy3,jy2);
1452             dz32             = _mm_sub_ps(iz3,jz2);
1453             dx33             = _mm_sub_ps(ix3,jx3);
1454             dy33             = _mm_sub_ps(iy3,jy3);
1455             dz33             = _mm_sub_ps(iz3,jz3);
1456
1457             /* Calculate squared distance and things based on it */
1458             rsq00            = gmx_mm_calc_rsq_ps(dx00,dy00,dz00);
1459             rsq11            = gmx_mm_calc_rsq_ps(dx11,dy11,dz11);
1460             rsq12            = gmx_mm_calc_rsq_ps(dx12,dy12,dz12);
1461             rsq13            = gmx_mm_calc_rsq_ps(dx13,dy13,dz13);
1462             rsq21            = gmx_mm_calc_rsq_ps(dx21,dy21,dz21);
1463             rsq22            = gmx_mm_calc_rsq_ps(dx22,dy22,dz22);
1464             rsq23            = gmx_mm_calc_rsq_ps(dx23,dy23,dz23);
1465             rsq31            = gmx_mm_calc_rsq_ps(dx31,dy31,dz31);
1466             rsq32            = gmx_mm_calc_rsq_ps(dx32,dy32,dz32);
1467             rsq33            = gmx_mm_calc_rsq_ps(dx33,dy33,dz33);
1468
1469             rinv11           = avx128fma_invsqrt_f(rsq11);
1470             rinv12           = avx128fma_invsqrt_f(rsq12);
1471             rinv13           = avx128fma_invsqrt_f(rsq13);
1472             rinv21           = avx128fma_invsqrt_f(rsq21);
1473             rinv22           = avx128fma_invsqrt_f(rsq22);
1474             rinv23           = avx128fma_invsqrt_f(rsq23);
1475             rinv31           = avx128fma_invsqrt_f(rsq31);
1476             rinv32           = avx128fma_invsqrt_f(rsq32);
1477             rinv33           = avx128fma_invsqrt_f(rsq33);
1478
1479             rinvsq00         = avx128fma_inv_f(rsq00);
1480             rinvsq11         = _mm_mul_ps(rinv11,rinv11);
1481             rinvsq12         = _mm_mul_ps(rinv12,rinv12);
1482             rinvsq13         = _mm_mul_ps(rinv13,rinv13);
1483             rinvsq21         = _mm_mul_ps(rinv21,rinv21);
1484             rinvsq22         = _mm_mul_ps(rinv22,rinv22);
1485             rinvsq23         = _mm_mul_ps(rinv23,rinv23);
1486             rinvsq31         = _mm_mul_ps(rinv31,rinv31);
1487             rinvsq32         = _mm_mul_ps(rinv32,rinv32);
1488             rinvsq33         = _mm_mul_ps(rinv33,rinv33);
1489
1490             fjx0             = _mm_setzero_ps();
1491             fjy0             = _mm_setzero_ps();
1492             fjz0             = _mm_setzero_ps();
1493             fjx1             = _mm_setzero_ps();
1494             fjy1             = _mm_setzero_ps();
1495             fjz1             = _mm_setzero_ps();
1496             fjx2             = _mm_setzero_ps();
1497             fjy2             = _mm_setzero_ps();
1498             fjz2             = _mm_setzero_ps();
1499             fjx3             = _mm_setzero_ps();
1500             fjy3             = _mm_setzero_ps();
1501             fjz3             = _mm_setzero_ps();
1502
1503             /**************************
1504              * CALCULATE INTERACTIONS *
1505              **************************/
1506
1507             /* LENNARD-JONES DISPERSION/REPULSION */
1508
1509             rinvsix          = _mm_mul_ps(_mm_mul_ps(rinvsq00,rinvsq00),rinvsq00);
1510             fvdw             = _mm_mul_ps(_mm_msub_ps(c12_00,rinvsix,c6_00),_mm_mul_ps(rinvsix,rinvsq00));
1511
1512             fscal            = fvdw;
1513
1514             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1515
1516              /* Update vectorial force */
1517             fix0             = _mm_macc_ps(dx00,fscal,fix0);
1518             fiy0             = _mm_macc_ps(dy00,fscal,fiy0);
1519             fiz0             = _mm_macc_ps(dz00,fscal,fiz0);
1520
1521             fjx0             = _mm_macc_ps(dx00,fscal,fjx0);
1522             fjy0             = _mm_macc_ps(dy00,fscal,fjy0);
1523             fjz0             = _mm_macc_ps(dz00,fscal,fjz0);
1524
1525             /**************************
1526              * CALCULATE INTERACTIONS *
1527              **************************/
1528
1529             /* COULOMB ELECTROSTATICS */
1530             velec            = _mm_mul_ps(qq11,rinv11);
1531             felec            = _mm_mul_ps(velec,rinvsq11);
1532
1533             fscal            = felec;
1534
1535             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1536
1537              /* Update vectorial force */
1538             fix1             = _mm_macc_ps(dx11,fscal,fix1);
1539             fiy1             = _mm_macc_ps(dy11,fscal,fiy1);
1540             fiz1             = _mm_macc_ps(dz11,fscal,fiz1);
1541
1542             fjx1             = _mm_macc_ps(dx11,fscal,fjx1);
1543             fjy1             = _mm_macc_ps(dy11,fscal,fjy1);
1544             fjz1             = _mm_macc_ps(dz11,fscal,fjz1);
1545
1546             /**************************
1547              * CALCULATE INTERACTIONS *
1548              **************************/
1549
1550             /* COULOMB ELECTROSTATICS */
1551             velec            = _mm_mul_ps(qq12,rinv12);
1552             felec            = _mm_mul_ps(velec,rinvsq12);
1553
1554             fscal            = felec;
1555
1556             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1557
1558              /* Update vectorial force */
1559             fix1             = _mm_macc_ps(dx12,fscal,fix1);
1560             fiy1             = _mm_macc_ps(dy12,fscal,fiy1);
1561             fiz1             = _mm_macc_ps(dz12,fscal,fiz1);
1562
1563             fjx2             = _mm_macc_ps(dx12,fscal,fjx2);
1564             fjy2             = _mm_macc_ps(dy12,fscal,fjy2);
1565             fjz2             = _mm_macc_ps(dz12,fscal,fjz2);
1566
1567             /**************************
1568              * CALCULATE INTERACTIONS *
1569              **************************/
1570
1571             /* COULOMB ELECTROSTATICS */
1572             velec            = _mm_mul_ps(qq13,rinv13);
1573             felec            = _mm_mul_ps(velec,rinvsq13);
1574
1575             fscal            = felec;
1576
1577             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1578
1579              /* Update vectorial force */
1580             fix1             = _mm_macc_ps(dx13,fscal,fix1);
1581             fiy1             = _mm_macc_ps(dy13,fscal,fiy1);
1582             fiz1             = _mm_macc_ps(dz13,fscal,fiz1);
1583
1584             fjx3             = _mm_macc_ps(dx13,fscal,fjx3);
1585             fjy3             = _mm_macc_ps(dy13,fscal,fjy3);
1586             fjz3             = _mm_macc_ps(dz13,fscal,fjz3);
1587
1588             /**************************
1589              * CALCULATE INTERACTIONS *
1590              **************************/
1591
1592             /* COULOMB ELECTROSTATICS */
1593             velec            = _mm_mul_ps(qq21,rinv21);
1594             felec            = _mm_mul_ps(velec,rinvsq21);
1595
1596             fscal            = felec;
1597
1598             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1599
1600              /* Update vectorial force */
1601             fix2             = _mm_macc_ps(dx21,fscal,fix2);
1602             fiy2             = _mm_macc_ps(dy21,fscal,fiy2);
1603             fiz2             = _mm_macc_ps(dz21,fscal,fiz2);
1604
1605             fjx1             = _mm_macc_ps(dx21,fscal,fjx1);
1606             fjy1             = _mm_macc_ps(dy21,fscal,fjy1);
1607             fjz1             = _mm_macc_ps(dz21,fscal,fjz1);
1608
1609             /**************************
1610              * CALCULATE INTERACTIONS *
1611              **************************/
1612
1613             /* COULOMB ELECTROSTATICS */
1614             velec            = _mm_mul_ps(qq22,rinv22);
1615             felec            = _mm_mul_ps(velec,rinvsq22);
1616
1617             fscal            = felec;
1618
1619             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1620
1621              /* Update vectorial force */
1622             fix2             = _mm_macc_ps(dx22,fscal,fix2);
1623             fiy2             = _mm_macc_ps(dy22,fscal,fiy2);
1624             fiz2             = _mm_macc_ps(dz22,fscal,fiz2);
1625
1626             fjx2             = _mm_macc_ps(dx22,fscal,fjx2);
1627             fjy2             = _mm_macc_ps(dy22,fscal,fjy2);
1628             fjz2             = _mm_macc_ps(dz22,fscal,fjz2);
1629
1630             /**************************
1631              * CALCULATE INTERACTIONS *
1632              **************************/
1633
1634             /* COULOMB ELECTROSTATICS */
1635             velec            = _mm_mul_ps(qq23,rinv23);
1636             felec            = _mm_mul_ps(velec,rinvsq23);
1637
1638             fscal            = felec;
1639
1640             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1641
1642              /* Update vectorial force */
1643             fix2             = _mm_macc_ps(dx23,fscal,fix2);
1644             fiy2             = _mm_macc_ps(dy23,fscal,fiy2);
1645             fiz2             = _mm_macc_ps(dz23,fscal,fiz2);
1646
1647             fjx3             = _mm_macc_ps(dx23,fscal,fjx3);
1648             fjy3             = _mm_macc_ps(dy23,fscal,fjy3);
1649             fjz3             = _mm_macc_ps(dz23,fscal,fjz3);
1650
1651             /**************************
1652              * CALCULATE INTERACTIONS *
1653              **************************/
1654
1655             /* COULOMB ELECTROSTATICS */
1656             velec            = _mm_mul_ps(qq31,rinv31);
1657             felec            = _mm_mul_ps(velec,rinvsq31);
1658
1659             fscal            = felec;
1660
1661             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1662
1663              /* Update vectorial force */
1664             fix3             = _mm_macc_ps(dx31,fscal,fix3);
1665             fiy3             = _mm_macc_ps(dy31,fscal,fiy3);
1666             fiz3             = _mm_macc_ps(dz31,fscal,fiz3);
1667
1668             fjx1             = _mm_macc_ps(dx31,fscal,fjx1);
1669             fjy1             = _mm_macc_ps(dy31,fscal,fjy1);
1670             fjz1             = _mm_macc_ps(dz31,fscal,fjz1);
1671
1672             /**************************
1673              * CALCULATE INTERACTIONS *
1674              **************************/
1675
1676             /* COULOMB ELECTROSTATICS */
1677             velec            = _mm_mul_ps(qq32,rinv32);
1678             felec            = _mm_mul_ps(velec,rinvsq32);
1679
1680             fscal            = felec;
1681
1682             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1683
1684              /* Update vectorial force */
1685             fix3             = _mm_macc_ps(dx32,fscal,fix3);
1686             fiy3             = _mm_macc_ps(dy32,fscal,fiy3);
1687             fiz3             = _mm_macc_ps(dz32,fscal,fiz3);
1688
1689             fjx2             = _mm_macc_ps(dx32,fscal,fjx2);
1690             fjy2             = _mm_macc_ps(dy32,fscal,fjy2);
1691             fjz2             = _mm_macc_ps(dz32,fscal,fjz2);
1692
1693             /**************************
1694              * CALCULATE INTERACTIONS *
1695              **************************/
1696
1697             /* COULOMB ELECTROSTATICS */
1698             velec            = _mm_mul_ps(qq33,rinv33);
1699             felec            = _mm_mul_ps(velec,rinvsq33);
1700
1701             fscal            = felec;
1702
1703             fscal            = _mm_andnot_ps(dummy_mask,fscal);
1704
1705              /* Update vectorial force */
1706             fix3             = _mm_macc_ps(dx33,fscal,fix3);
1707             fiy3             = _mm_macc_ps(dy33,fscal,fiy3);
1708             fiz3             = _mm_macc_ps(dz33,fscal,fiz3);
1709
1710             fjx3             = _mm_macc_ps(dx33,fscal,fjx3);
1711             fjy3             = _mm_macc_ps(dy33,fscal,fjy3);
1712             fjz3             = _mm_macc_ps(dz33,fscal,fjz3);
1713
1714             fjptrA             = (jnrlistA>=0) ? f+j_coord_offsetA : scratch;
1715             fjptrB             = (jnrlistB>=0) ? f+j_coord_offsetB : scratch;
1716             fjptrC             = (jnrlistC>=0) ? f+j_coord_offsetC : scratch;
1717             fjptrD             = (jnrlistD>=0) ? f+j_coord_offsetD : scratch;
1718
1719             gmx_mm_decrement_4rvec_4ptr_swizzle_ps(fjptrA,fjptrB,fjptrC,fjptrD,
1720                                                    fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,
1721                                                    fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1722
1723             /* Inner loop uses 303 flops */
1724         }
1725
1726         /* End of innermost loop */
1727
1728         gmx_mm_update_iforce_4atom_swizzle_ps(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1729                                               f+i_coord_offset,fshift+i_shift_offset);
1730
1731         /* Increment number of inner iterations */
1732         inneriter                  += j_index_end - j_index_start;
1733
1734         /* Outer loop uses 24 flops */
1735     }
1736
1737     /* Increment number of outer iterations */
1738     outeriter        += nri;
1739
1740     /* Update outer/inner flops */
1741
1742     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*303);
1743 }