src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecRF_VdwLJ_GeomW4W4_avx_128_fma_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 /*
  36  * Note: this file was generated by the GROMACS avx_128_fma_double kernel generator.
  37  */
  38 #include "gmxpre.h"
  39
  40 #include "config.h"
  41
  42 #include <math.h>
  43
  44 #include "../nb_kernel.h"
  45 #include "gromacs/legacyheaders/types/simple.h"
  46 #include "gromacs/math/vec.h"
  47 #include "gromacs/legacyheaders/nrnb.h"
  48
  49 #include "gromacs/simd/math_x86_avx_128_fma_double.h"
  50 #include "kernelutil_x86_avx_128_fma_double.h"
  51
  52 /*
  53  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW4W4_VF_avx_128_fma_double
  54  * Electrostatics interaction: ReactionField
  55  * VdW interaction:            LennardJones
  56  * Geometry:                   Water4-Water4
  57  * Calculate force/pot:        PotentialAndForce
  58  */
  59 void
  60 nb_kernel_ElecRF_VdwLJ_GeomW4W4_VF_avx_128_fma_double
  61                     (t_nblist                    * gmx_restrict       nlist,
  62                      rvec                        * gmx_restrict          xx,
  63                      rvec                        * gmx_restrict          ff,
  64                      t_forcerec                  * gmx_restrict          fr,
  65                      t_mdatoms                   * gmx_restrict     mdatoms,
  66                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  67                      t_nrnb                      * gmx_restrict        nrnb)
  68 {
  69     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  70      * just 0 for non-waters.
  71      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  72      * jnr indices corresponding to data put in the four positions in the SIMD register.
  73      */
  74     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  75     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  76     int              jnrA,jnrB;
  77     int              j_coord_offsetA,j_coord_offsetB;
  78     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  79     real             rcutoff_scalar;
  80     real             *shiftvec,*fshift,*x,*f;
  81     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  82     int              vdwioffset0;
  83     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  84     int              vdwioffset1;
  85     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  86     int              vdwioffset2;
  87     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  88     int              vdwioffset3;
  89     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  90     int              vdwjidx0A,vdwjidx0B;
  91     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  92     int              vdwjidx1A,vdwjidx1B;
  93     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  94     int              vdwjidx2A,vdwjidx2B;
  95     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  96     int              vdwjidx3A,vdwjidx3B;
  97     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  98     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  99     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 100     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 101     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 102     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 103     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 104     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 105     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 106     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 107     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 108     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 109     real             *charge;
 110     int              nvdwtype;
 111     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 112     int              *vdwtype;
 113     real             *vdwparam;
 114     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 115     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 116     __m128d          dummy_mask,cutoff_mask;
 117     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 118     __m128d          one     = _mm_set1_pd(1.0);
 119     __m128d          two     = _mm_set1_pd(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_pd(fr->epsfac);
 132     charge           = mdatoms->chargeA;
 133     krf              = _mm_set1_pd(fr->ic->k_rf);
 134     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 135     crf              = _mm_set1_pd(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_pd(facel,_mm_set1_pd(charge[inr+1]));
 143     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 144     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 145     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 146
 147     jq1              = _mm_set1_pd(charge[inr+1]);
 148     jq2              = _mm_set1_pd(charge[inr+2]);
 149     jq3              = _mm_set1_pd(charge[inr+3]);
 150     vdwjidx0A        = 2*vdwtype[inr+0];
 151     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 152     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 153     qq11             = _mm_mul_pd(iq1,jq1);
 154     qq12             = _mm_mul_pd(iq1,jq2);
 155     qq13             = _mm_mul_pd(iq1,jq3);
 156     qq21             = _mm_mul_pd(iq2,jq1);
 157     qq22             = _mm_mul_pd(iq2,jq2);
 158     qq23             = _mm_mul_pd(iq2,jq3);
 159     qq31             = _mm_mul_pd(iq3,jq1);
 160     qq32             = _mm_mul_pd(iq3,jq2);
 161     qq33             = _mm_mul_pd(iq3,jq3);
 162
 163     /* Avoid stupid compiler warnings */
 164     jnrA = jnrB = 0;
 165     j_coord_offsetA = 0;
 166     j_coord_offsetB = 0;
 167
 168     outeriter        = 0;
 169     inneriter        = 0;
 170
 171     /* Start outer loop over neighborlists */
 172     for(iidx=0; iidx<nri; iidx++)
 173     {
 174         /* Load shift vector for this list */
 175         i_shift_offset   = DIM*shiftidx[iidx];
 176
 177         /* Load limits for loop over neighbors */
 178         j_index_start    = jindex[iidx];
 179         j_index_end      = jindex[iidx+1];
 180
 181         /* Get outer coordinate index */
 182         inr              = iinr[iidx];
 183         i_coord_offset   = DIM*inr;
 184
 185         /* Load i particle coords and add shift vector */
 186         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 187                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 188
 189         fix0             = _mm_setzero_pd();
 190         fiy0             = _mm_setzero_pd();
 191         fiz0             = _mm_setzero_pd();
 192         fix1             = _mm_setzero_pd();
 193         fiy1             = _mm_setzero_pd();
 194         fiz1             = _mm_setzero_pd();
 195         fix2             = _mm_setzero_pd();
 196         fiy2             = _mm_setzero_pd();
 197         fiz2             = _mm_setzero_pd();
 198         fix3             = _mm_setzero_pd();
 199         fiy3             = _mm_setzero_pd();
 200         fiz3             = _mm_setzero_pd();
 201
 202         /* Reset potential sums */
 203         velecsum         = _mm_setzero_pd();
 204         vvdwsum          = _mm_setzero_pd();
 205
 206         /* Start inner kernel loop */
 207         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 208         {
 209
 210             /* Get j neighbor index, and coordinate index */
 211             jnrA             = jjnr[jidx];
 212             jnrB             = jjnr[jidx+1];
 213             j_coord_offsetA  = DIM*jnrA;
 214             j_coord_offsetB  = DIM*jnrB;
 215
 216             /* load j atom coordinates */
 217             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 218                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 219                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 220
 221             /* Calculate displacement vector */
 222             dx00             = _mm_sub_pd(ix0,jx0);
 223             dy00             = _mm_sub_pd(iy0,jy0);
 224             dz00             = _mm_sub_pd(iz0,jz0);
 225             dx11             = _mm_sub_pd(ix1,jx1);
 226             dy11             = _mm_sub_pd(iy1,jy1);
 227             dz11             = _mm_sub_pd(iz1,jz1);
 228             dx12             = _mm_sub_pd(ix1,jx2);
 229             dy12             = _mm_sub_pd(iy1,jy2);
 230             dz12             = _mm_sub_pd(iz1,jz2);
 231             dx13             = _mm_sub_pd(ix1,jx3);
 232             dy13             = _mm_sub_pd(iy1,jy3);
 233             dz13             = _mm_sub_pd(iz1,jz3);
 234             dx21             = _mm_sub_pd(ix2,jx1);
 235             dy21             = _mm_sub_pd(iy2,jy1);
 236             dz21             = _mm_sub_pd(iz2,jz1);
 237             dx22             = _mm_sub_pd(ix2,jx2);
 238             dy22             = _mm_sub_pd(iy2,jy2);
 239             dz22             = _mm_sub_pd(iz2,jz2);
 240             dx23             = _mm_sub_pd(ix2,jx3);
 241             dy23             = _mm_sub_pd(iy2,jy3);
 242             dz23             = _mm_sub_pd(iz2,jz3);
 243             dx31             = _mm_sub_pd(ix3,jx1);
 244             dy31             = _mm_sub_pd(iy3,jy1);
 245             dz31             = _mm_sub_pd(iz3,jz1);
 246             dx32             = _mm_sub_pd(ix3,jx2);
 247             dy32             = _mm_sub_pd(iy3,jy2);
 248             dz32             = _mm_sub_pd(iz3,jz2);
 249             dx33             = _mm_sub_pd(ix3,jx3);
 250             dy33             = _mm_sub_pd(iy3,jy3);
 251             dz33             = _mm_sub_pd(iz3,jz3);
 252
 253             /* Calculate squared distance and things based on it */
 254             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 255             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 256             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 257             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 258             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 259             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 260             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 261             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 262             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 263             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 264
 265             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 266             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 267             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 268             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 269             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 270             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 271             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 272             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 273             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 274
 275             rinvsq00         = gmx_mm_inv_pd(rsq00);
 276             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 277             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 278             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 279             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 280             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 281             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 282             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 283             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 284             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 285
 286             fjx0             = _mm_setzero_pd();
 287             fjy0             = _mm_setzero_pd();
 288             fjz0             = _mm_setzero_pd();
 289             fjx1             = _mm_setzero_pd();
 290             fjy1             = _mm_setzero_pd();
 291             fjz1             = _mm_setzero_pd();
 292             fjx2             = _mm_setzero_pd();
 293             fjy2             = _mm_setzero_pd();
 294             fjz2             = _mm_setzero_pd();
 295             fjx3             = _mm_setzero_pd();
 296             fjy3             = _mm_setzero_pd();
 297             fjz3             = _mm_setzero_pd();
 298
 299             /**************************
 300              * CALCULATE INTERACTIONS *
 301              **************************/
 302
 303             /* LENNARD-JONES DISPERSION/REPULSION */
 304
 305             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 306             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 307             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 308             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 309             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 310
 311             /* Update potential sum for this i atom from the interaction with this j atom. */
 312             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 313
 314             fscal            = fvdw;
 315
 316             /* Update vectorial force */
 317             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 318             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 319             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 320
 321             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 322             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 323             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 324
 325             /**************************
 326              * CALCULATE INTERACTIONS *
 327              **************************/
 328
 329             /* REACTION-FIELD ELECTROSTATICS */
 330             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 331             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 332
 333             /* Update potential sum for this i atom from the interaction with this j atom. */
 334             velecsum         = _mm_add_pd(velecsum,velec);
 335
 336             fscal            = felec;
 337
 338             /* Update vectorial force */
 339             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 340             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 341             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 342
 343             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 344             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 345             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 346
 347             /**************************
 348              * CALCULATE INTERACTIONS *
 349              **************************/
 350
 351             /* REACTION-FIELD ELECTROSTATICS */
 352             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 353             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 354
 355             /* Update potential sum for this i atom from the interaction with this j atom. */
 356             velecsum         = _mm_add_pd(velecsum,velec);
 357
 358             fscal            = felec;
 359
 360             /* Update vectorial force */
 361             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 362             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 363             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 364
 365             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 366             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 367             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 368
 369             /**************************
 370              * CALCULATE INTERACTIONS *
 371              **************************/
 372
 373             /* REACTION-FIELD ELECTROSTATICS */
 374             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 375             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 376
 377             /* Update potential sum for this i atom from the interaction with this j atom. */
 378             velecsum         = _mm_add_pd(velecsum,velec);
 379
 380             fscal            = felec;
 381
 382             /* Update vectorial force */
 383             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 384             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 385             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 386
 387             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 388             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 389             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 390
 391             /**************************
 392              * CALCULATE INTERACTIONS *
 393              **************************/
 394
 395             /* REACTION-FIELD ELECTROSTATICS */
 396             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 397             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 398
 399             /* Update potential sum for this i atom from the interaction with this j atom. */
 400             velecsum         = _mm_add_pd(velecsum,velec);
 401
 402             fscal            = felec;
 403
 404             /* Update vectorial force */
 405             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 406             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 407             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 408
 409             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 410             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 411             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 412
 413             /**************************
 414              * CALCULATE INTERACTIONS *
 415              **************************/
 416
 417             /* REACTION-FIELD ELECTROSTATICS */
 418             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 419             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 420
 421             /* Update potential sum for this i atom from the interaction with this j atom. */
 422             velecsum         = _mm_add_pd(velecsum,velec);
 423
 424             fscal            = felec;
 425
 426             /* Update vectorial force */
 427             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 428             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 429             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 430
 431             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 432             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 433             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 434
 435             /**************************
 436              * CALCULATE INTERACTIONS *
 437              **************************/
 438
 439             /* REACTION-FIELD ELECTROSTATICS */
 440             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 441             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 442
 443             /* Update potential sum for this i atom from the interaction with this j atom. */
 444             velecsum         = _mm_add_pd(velecsum,velec);
 445
 446             fscal            = felec;
 447
 448             /* Update vectorial force */
 449             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 450             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 451             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 452
 453             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 454             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 455             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 456
 457             /**************************
 458              * CALCULATE INTERACTIONS *
 459              **************************/
 460
 461             /* REACTION-FIELD ELECTROSTATICS */
 462             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
 463             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
 464
 465             /* Update potential sum for this i atom from the interaction with this j atom. */
 466             velecsum         = _mm_add_pd(velecsum,velec);
 467
 468             fscal            = felec;
 469
 470             /* Update vectorial force */
 471             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 472             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 473             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 474
 475             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 476             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 477             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 478
 479             /**************************
 480              * CALCULATE INTERACTIONS *
 481              **************************/
 482
 483             /* REACTION-FIELD ELECTROSTATICS */
 484             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
 485             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
 486
 487             /* Update potential sum for this i atom from the interaction with this j atom. */
 488             velecsum         = _mm_add_pd(velecsum,velec);
 489
 490             fscal            = felec;
 491
 492             /* Update vectorial force */
 493             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 494             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 495             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 496
 497             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 498             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 499             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 500
 501             /**************************
 502              * CALCULATE INTERACTIONS *
 503              **************************/
 504
 505             /* REACTION-FIELD ELECTROSTATICS */
 506             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
 507             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
 508
 509             /* Update potential sum for this i atom from the interaction with this j atom. */
 510             velecsum         = _mm_add_pd(velecsum,velec);
 511
 512             fscal            = felec;
 513
 514             /* Update vectorial force */
 515             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 516             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 517             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 518
 519             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 520             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 521             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 522
 523             gmx_mm_decrement_4rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 524
 525             /* Inner loop uses 353 flops */
 526         }
 527
 528         if(jidx<j_index_end)
 529         {
 530
 531             jnrA             = jjnr[jidx];
 532             j_coord_offsetA  = DIM*jnrA;
 533
 534             /* load j atom coordinates */
 535             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 536                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 537                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 538
 539             /* Calculate displacement vector */
 540             dx00             = _mm_sub_pd(ix0,jx0);
 541             dy00             = _mm_sub_pd(iy0,jy0);
 542             dz00             = _mm_sub_pd(iz0,jz0);
 543             dx11             = _mm_sub_pd(ix1,jx1);
 544             dy11             = _mm_sub_pd(iy1,jy1);
 545             dz11             = _mm_sub_pd(iz1,jz1);
 546             dx12             = _mm_sub_pd(ix1,jx2);
 547             dy12             = _mm_sub_pd(iy1,jy2);
 548             dz12             = _mm_sub_pd(iz1,jz2);
 549             dx13             = _mm_sub_pd(ix1,jx3);
 550             dy13             = _mm_sub_pd(iy1,jy3);
 551             dz13             = _mm_sub_pd(iz1,jz3);
 552             dx21             = _mm_sub_pd(ix2,jx1);
 553             dy21             = _mm_sub_pd(iy2,jy1);
 554             dz21             = _mm_sub_pd(iz2,jz1);
 555             dx22             = _mm_sub_pd(ix2,jx2);
 556             dy22             = _mm_sub_pd(iy2,jy2);
 557             dz22             = _mm_sub_pd(iz2,jz2);
 558             dx23             = _mm_sub_pd(ix2,jx3);
 559             dy23             = _mm_sub_pd(iy2,jy3);
 560             dz23             = _mm_sub_pd(iz2,jz3);
 561             dx31             = _mm_sub_pd(ix3,jx1);
 562             dy31             = _mm_sub_pd(iy3,jy1);
 563             dz31             = _mm_sub_pd(iz3,jz1);
 564             dx32             = _mm_sub_pd(ix3,jx2);
 565             dy32             = _mm_sub_pd(iy3,jy2);
 566             dz32             = _mm_sub_pd(iz3,jz2);
 567             dx33             = _mm_sub_pd(ix3,jx3);
 568             dy33             = _mm_sub_pd(iy3,jy3);
 569             dz33             = _mm_sub_pd(iz3,jz3);
 570
 571             /* Calculate squared distance and things based on it */
 572             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 573             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 574             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 575             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 576             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 577             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 578             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 579             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 580             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 581             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 582
 583             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 584             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 585             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 586             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 587             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 588             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 589             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 590             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 591             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 592
 593             rinvsq00         = gmx_mm_inv_pd(rsq00);
 594             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 595             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 596             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 597             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 598             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 599             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 600             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 601             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 602             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 603
 604             fjx0             = _mm_setzero_pd();
 605             fjy0             = _mm_setzero_pd();
 606             fjz0             = _mm_setzero_pd();
 607             fjx1             = _mm_setzero_pd();
 608             fjy1             = _mm_setzero_pd();
 609             fjz1             = _mm_setzero_pd();
 610             fjx2             = _mm_setzero_pd();
 611             fjy2             = _mm_setzero_pd();
 612             fjz2             = _mm_setzero_pd();
 613             fjx3             = _mm_setzero_pd();
 614             fjy3             = _mm_setzero_pd();
 615             fjz3             = _mm_setzero_pd();
 616
 617             /**************************
 618              * CALCULATE INTERACTIONS *
 619              **************************/
 620
 621             /* LENNARD-JONES DISPERSION/REPULSION */
 622
 623             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 624             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 625             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 626             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 627             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 628
 629             /* Update potential sum for this i atom from the interaction with this j atom. */
 630             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 631             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 632
 633             fscal            = fvdw;
 634
 635             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 636
 637             /* Update vectorial force */
 638             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 639             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 640             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 641
 642             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 643             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 644             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 645
 646             /**************************
 647              * CALCULATE INTERACTIONS *
 648              **************************/
 649
 650             /* REACTION-FIELD ELECTROSTATICS */
 651             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 652             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 653
 654             /* Update potential sum for this i atom from the interaction with this j atom. */
 655             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 656             velecsum         = _mm_add_pd(velecsum,velec);
 657
 658             fscal            = felec;
 659
 660             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 661
 662             /* Update vectorial force */
 663             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 664             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 665             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 666
 667             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 668             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 669             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 670
 671             /**************************
 672              * CALCULATE INTERACTIONS *
 673              **************************/
 674
 675             /* REACTION-FIELD ELECTROSTATICS */
 676             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 677             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 678
 679             /* Update potential sum for this i atom from the interaction with this j atom. */
 680             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 681             velecsum         = _mm_add_pd(velecsum,velec);
 682
 683             fscal            = felec;
 684
 685             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 686
 687             /* Update vectorial force */
 688             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 689             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 690             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 691
 692             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 693             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 694             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 695
 696             /**************************
 697              * CALCULATE INTERACTIONS *
 698              **************************/
 699
 700             /* REACTION-FIELD ELECTROSTATICS */
 701             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 702             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 703
 704             /* Update potential sum for this i atom from the interaction with this j atom. */
 705             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 706             velecsum         = _mm_add_pd(velecsum,velec);
 707
 708             fscal            = felec;
 709
 710             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 711
 712             /* Update vectorial force */
 713             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 714             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 715             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 716
 717             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 718             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 719             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 720
 721             /**************************
 722              * CALCULATE INTERACTIONS *
 723              **************************/
 724
 725             /* REACTION-FIELD ELECTROSTATICS */
 726             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 727             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 728
 729             /* Update potential sum for this i atom from the interaction with this j atom. */
 730             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 731             velecsum         = _mm_add_pd(velecsum,velec);
 732
 733             fscal            = felec;
 734
 735             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 736
 737             /* Update vectorial force */
 738             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 739             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 740             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 741
 742             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 743             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 744             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 745
 746             /**************************
 747              * CALCULATE INTERACTIONS *
 748              **************************/
 749
 750             /* REACTION-FIELD ELECTROSTATICS */
 751             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 752             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 753
 754             /* Update potential sum for this i atom from the interaction with this j atom. */
 755             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 756             velecsum         = _mm_add_pd(velecsum,velec);
 757
 758             fscal            = felec;
 759
 760             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 761
 762             /* Update vectorial force */
 763             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 764             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 765             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 766
 767             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 768             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 769             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 770
 771             /**************************
 772              * CALCULATE INTERACTIONS *
 773              **************************/
 774
 775             /* REACTION-FIELD ELECTROSTATICS */
 776             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 777             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 778
 779             /* Update potential sum for this i atom from the interaction with this j atom. */
 780             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 781             velecsum         = _mm_add_pd(velecsum,velec);
 782
 783             fscal            = felec;
 784
 785             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 786
 787             /* Update vectorial force */
 788             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 789             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 790             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 791
 792             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 793             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 794             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 795
 796             /**************************
 797              * CALCULATE INTERACTIONS *
 798              **************************/
 799
 800             /* REACTION-FIELD ELECTROSTATICS */
 801             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
 802             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
 803
 804             /* Update potential sum for this i atom from the interaction with this j atom. */
 805             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 806             velecsum         = _mm_add_pd(velecsum,velec);
 807
 808             fscal            = felec;
 809
 810             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 811
 812             /* Update vectorial force */
 813             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 814             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 815             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 816
 817             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 818             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 819             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 820
 821             /**************************
 822              * CALCULATE INTERACTIONS *
 823              **************************/
 824
 825             /* REACTION-FIELD ELECTROSTATICS */
 826             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
 827             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
 828
 829             /* Update potential sum for this i atom from the interaction with this j atom. */
 830             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 831             velecsum         = _mm_add_pd(velecsum,velec);
 832
 833             fscal            = felec;
 834
 835             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 836
 837             /* Update vectorial force */
 838             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 839             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 840             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 841
 842             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 843             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 844             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 845
 846             /**************************
 847              * CALCULATE INTERACTIONS *
 848              **************************/
 849
 850             /* REACTION-FIELD ELECTROSTATICS */
 851             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
 852             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
 853
 854             /* Update potential sum for this i atom from the interaction with this j atom. */
 855             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 856             velecsum         = _mm_add_pd(velecsum,velec);
 857
 858             fscal            = felec;
 859
 860             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 861
 862             /* Update vectorial force */
 863             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 864             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 865             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 866
 867             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 868             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 869             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 870
 871             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 872
 873             /* Inner loop uses 353 flops */
 874         }
 875
 876         /* End of innermost loop */
 877
 878         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 879                                               f+i_coord_offset,fshift+i_shift_offset);
 880
 881         ggid                        = gid[iidx];
 882         /* Update potential energies */
 883         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 884         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 885
 886         /* Increment number of inner iterations */
 887         inneriter                  += j_index_end - j_index_start;
 888
 889         /* Outer loop uses 26 flops */
 890     }
 891
 892     /* Increment number of outer iterations */
 893     outeriter        += nri;
 894
 895     /* Update outer/inner flops */
 896
 897     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*353);
 898 }
 899 /*
 900  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW4W4_F_avx_128_fma_double
 901  * Electrostatics interaction: ReactionField
 902  * VdW interaction:            LennardJones
 903  * Geometry:                   Water4-Water4
 904  * Calculate force/pot:        Force
 905  */
 906 void
 907 nb_kernel_ElecRF_VdwLJ_GeomW4W4_F_avx_128_fma_double
 908                     (t_nblist                    * gmx_restrict       nlist,
 909                      rvec                        * gmx_restrict          xx,
 910                      rvec                        * gmx_restrict          ff,
 911                      t_forcerec                  * gmx_restrict          fr,
 912                      t_mdatoms                   * gmx_restrict     mdatoms,
 913                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 914                      t_nrnb                      * gmx_restrict        nrnb)
 915 {
 916     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 917      * just 0 for non-waters.
 918      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 919      * jnr indices corresponding to data put in the four positions in the SIMD register.
 920      */
 921     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 922     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 923     int              jnrA,jnrB;
 924     int              j_coord_offsetA,j_coord_offsetB;
 925     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 926     real             rcutoff_scalar;
 927     real             *shiftvec,*fshift,*x,*f;
 928     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 929     int              vdwioffset0;
 930     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 931     int              vdwioffset1;
 932     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 933     int              vdwioffset2;
 934     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 935     int              vdwioffset3;
 936     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 937     int              vdwjidx0A,vdwjidx0B;
 938     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 939     int              vdwjidx1A,vdwjidx1B;
 940     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 941     int              vdwjidx2A,vdwjidx2B;
 942     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 943     int              vdwjidx3A,vdwjidx3B;
 944     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 945     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 946     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 947     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 948     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 949     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 950     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 951     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 952     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 953     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 954     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 955     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 956     real             *charge;
 957     int              nvdwtype;
 958     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 959     int              *vdwtype;
 960     real             *vdwparam;
 961     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 962     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 963     __m128d          dummy_mask,cutoff_mask;
 964     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 965     __m128d          one     = _mm_set1_pd(1.0);
 966     __m128d          two     = _mm_set1_pd(2.0);
 967     x                = xx[0];
 968     f                = ff[0];
 969
 970     nri              = nlist->nri;
 971     iinr             = nlist->iinr;
 972     jindex           = nlist->jindex;
 973     jjnr             = nlist->jjnr;
 974     shiftidx         = nlist->shift;
 975     gid              = nlist->gid;
 976     shiftvec         = fr->shift_vec[0];
 977     fshift           = fr->fshift[0];
 978     facel            = _mm_set1_pd(fr->epsfac);
 979     charge           = mdatoms->chargeA;
 980     krf              = _mm_set1_pd(fr->ic->k_rf);
 981     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 982     crf              = _mm_set1_pd(fr->ic->c_rf);
 983     nvdwtype         = fr->ntype;
 984     vdwparam         = fr->nbfp;
 985     vdwtype          = mdatoms->typeA;
 986
 987     /* Setup water-specific parameters */
 988     inr              = nlist->iinr[0];
 989     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 990     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 991     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 992     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 993
 994     jq1              = _mm_set1_pd(charge[inr+1]);
 995     jq2              = _mm_set1_pd(charge[inr+2]);
 996     jq3              = _mm_set1_pd(charge[inr+3]);
 997     vdwjidx0A        = 2*vdwtype[inr+0];
 998     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 999     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1000     qq11             = _mm_mul_pd(iq1,jq1);
1001     qq12             = _mm_mul_pd(iq1,jq2);
1002     qq13             = _mm_mul_pd(iq1,jq3);
1003     qq21             = _mm_mul_pd(iq2,jq1);
1004     qq22             = _mm_mul_pd(iq2,jq2);
1005     qq23             = _mm_mul_pd(iq2,jq3);
1006     qq31             = _mm_mul_pd(iq3,jq1);
1007     qq32             = _mm_mul_pd(iq3,jq2);
1008     qq33             = _mm_mul_pd(iq3,jq3);
1009
1010     /* Avoid stupid compiler warnings */
1011     jnrA = jnrB = 0;
1012     j_coord_offsetA = 0;
1013     j_coord_offsetB = 0;
1014
1015     outeriter        = 0;
1016     inneriter        = 0;
1017
1018     /* Start outer loop over neighborlists */
1019     for(iidx=0; iidx<nri; iidx++)
1020     {
1021         /* Load shift vector for this list */
1022         i_shift_offset   = DIM*shiftidx[iidx];
1023
1024         /* Load limits for loop over neighbors */
1025         j_index_start    = jindex[iidx];
1026         j_index_end      = jindex[iidx+1];
1027
1028         /* Get outer coordinate index */
1029         inr              = iinr[iidx];
1030         i_coord_offset   = DIM*inr;
1031
1032         /* Load i particle coords and add shift vector */
1033         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1034                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1035
1036         fix0             = _mm_setzero_pd();
1037         fiy0             = _mm_setzero_pd();
1038         fiz0             = _mm_setzero_pd();
1039         fix1             = _mm_setzero_pd();
1040         fiy1             = _mm_setzero_pd();
1041         fiz1             = _mm_setzero_pd();
1042         fix2             = _mm_setzero_pd();
1043         fiy2             = _mm_setzero_pd();
1044         fiz2             = _mm_setzero_pd();
1045         fix3             = _mm_setzero_pd();
1046         fiy3             = _mm_setzero_pd();
1047         fiz3             = _mm_setzero_pd();
1048
1049         /* Start inner kernel loop */
1050         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1051         {
1052
1053             /* Get j neighbor index, and coordinate index */
1054             jnrA             = jjnr[jidx];
1055             jnrB             = jjnr[jidx+1];
1056             j_coord_offsetA  = DIM*jnrA;
1057             j_coord_offsetB  = DIM*jnrB;
1058
1059             /* load j atom coordinates */
1060             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1061                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1062                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1063
1064             /* Calculate displacement vector */
1065             dx00             = _mm_sub_pd(ix0,jx0);
1066             dy00             = _mm_sub_pd(iy0,jy0);
1067             dz00             = _mm_sub_pd(iz0,jz0);
1068             dx11             = _mm_sub_pd(ix1,jx1);
1069             dy11             = _mm_sub_pd(iy1,jy1);
1070             dz11             = _mm_sub_pd(iz1,jz1);
1071             dx12             = _mm_sub_pd(ix1,jx2);
1072             dy12             = _mm_sub_pd(iy1,jy2);
1073             dz12             = _mm_sub_pd(iz1,jz2);
1074             dx13             = _mm_sub_pd(ix1,jx3);
1075             dy13             = _mm_sub_pd(iy1,jy3);
1076             dz13             = _mm_sub_pd(iz1,jz3);
1077             dx21             = _mm_sub_pd(ix2,jx1);
1078             dy21             = _mm_sub_pd(iy2,jy1);
1079             dz21             = _mm_sub_pd(iz2,jz1);
1080             dx22             = _mm_sub_pd(ix2,jx2);
1081             dy22             = _mm_sub_pd(iy2,jy2);
1082             dz22             = _mm_sub_pd(iz2,jz2);
1083             dx23             = _mm_sub_pd(ix2,jx3);
1084             dy23             = _mm_sub_pd(iy2,jy3);
1085             dz23             = _mm_sub_pd(iz2,jz3);
1086             dx31             = _mm_sub_pd(ix3,jx1);
1087             dy31             = _mm_sub_pd(iy3,jy1);
1088             dz31             = _mm_sub_pd(iz3,jz1);
1089             dx32             = _mm_sub_pd(ix3,jx2);
1090             dy32             = _mm_sub_pd(iy3,jy2);
1091             dz32             = _mm_sub_pd(iz3,jz2);
1092             dx33             = _mm_sub_pd(ix3,jx3);
1093             dy33             = _mm_sub_pd(iy3,jy3);
1094             dz33             = _mm_sub_pd(iz3,jz3);
1095
1096             /* Calculate squared distance and things based on it */
1097             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1098             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1099             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1100             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1101             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1102             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1103             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1104             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1105             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1106             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1107
1108             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1109             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1110             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1111             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1112             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1113             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1114             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1115             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1116             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1117
1118             rinvsq00         = gmx_mm_inv_pd(rsq00);
1119             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1120             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1121             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1122             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1123             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1124             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1125             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1126             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1127             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1128
1129             fjx0             = _mm_setzero_pd();
1130             fjy0             = _mm_setzero_pd();
1131             fjz0             = _mm_setzero_pd();
1132             fjx1             = _mm_setzero_pd();
1133             fjy1             = _mm_setzero_pd();
1134             fjz1             = _mm_setzero_pd();
1135             fjx2             = _mm_setzero_pd();
1136             fjy2             = _mm_setzero_pd();
1137             fjz2             = _mm_setzero_pd();
1138             fjx3             = _mm_setzero_pd();
1139             fjy3             = _mm_setzero_pd();
1140             fjz3             = _mm_setzero_pd();
1141
1142             /**************************
1143              * CALCULATE INTERACTIONS *
1144              **************************/
1145
1146             /* LENNARD-JONES DISPERSION/REPULSION */
1147
1148             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1149             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1150
1151             fscal            = fvdw;
1152
1153             /* Update vectorial force */
1154             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1155             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1156             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1157
1158             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1159             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1160             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1161
1162             /**************************
1163              * CALCULATE INTERACTIONS *
1164              **************************/
1165
1166             /* REACTION-FIELD ELECTROSTATICS */
1167             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1168
1169             fscal            = felec;
1170
1171             /* Update vectorial force */
1172             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1173             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1174             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1175
1176             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1177             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1178             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1179
1180             /**************************
1181              * CALCULATE INTERACTIONS *
1182              **************************/
1183
1184             /* REACTION-FIELD ELECTROSTATICS */
1185             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1186
1187             fscal            = felec;
1188
1189             /* Update vectorial force */
1190             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1191             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1192             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1193
1194             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1195             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1196             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1197
1198             /**************************
1199              * CALCULATE INTERACTIONS *
1200              **************************/
1201
1202             /* REACTION-FIELD ELECTROSTATICS */
1203             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1204
1205             fscal            = felec;
1206
1207             /* Update vectorial force */
1208             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1209             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1210             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1211
1212             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1213             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1214             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1215
1216             /**************************
1217              * CALCULATE INTERACTIONS *
1218              **************************/
1219
1220             /* REACTION-FIELD ELECTROSTATICS */
1221             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1222
1223             fscal            = felec;
1224
1225             /* Update vectorial force */
1226             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1227             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1228             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1229
1230             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1231             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1232             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1233
1234             /**************************
1235              * CALCULATE INTERACTIONS *
1236              **************************/
1237
1238             /* REACTION-FIELD ELECTROSTATICS */
1239             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1240
1241             fscal            = felec;
1242
1243             /* Update vectorial force */
1244             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1245             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1246             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1247
1248             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1249             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1250             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1251
1252             /**************************
1253              * CALCULATE INTERACTIONS *
1254              **************************/
1255
1256             /* REACTION-FIELD ELECTROSTATICS */
1257             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1258
1259             fscal            = felec;
1260
1261             /* Update vectorial force */
1262             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1263             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1264             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1265
1266             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1267             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1268             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1269
1270             /**************************
1271              * CALCULATE INTERACTIONS *
1272              **************************/
1273
1274             /* REACTION-FIELD ELECTROSTATICS */
1275             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1276
1277             fscal            = felec;
1278
1279             /* Update vectorial force */
1280             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1281             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1282             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1283
1284             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1285             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1286             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1287
1288             /**************************
1289              * CALCULATE INTERACTIONS *
1290              **************************/
1291
1292             /* REACTION-FIELD ELECTROSTATICS */
1293             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1294
1295             fscal            = felec;
1296
1297             /* Update vectorial force */
1298             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1299             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1300             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1301
1302             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1303             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1304             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1305
1306             /**************************
1307              * CALCULATE INTERACTIONS *
1308              **************************/
1309
1310             /* REACTION-FIELD ELECTROSTATICS */
1311             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1312
1313             fscal            = felec;
1314
1315             /* Update vectorial force */
1316             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1317             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1318             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1319
1320             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1321             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1322             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1323
1324             gmx_mm_decrement_4rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1325
1326             /* Inner loop uses 303 flops */
1327         }
1328
1329         if(jidx<j_index_end)
1330         {
1331
1332             jnrA             = jjnr[jidx];
1333             j_coord_offsetA  = DIM*jnrA;
1334
1335             /* load j atom coordinates */
1336             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1337                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1338                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1339
1340             /* Calculate displacement vector */
1341             dx00             = _mm_sub_pd(ix0,jx0);
1342             dy00             = _mm_sub_pd(iy0,jy0);
1343             dz00             = _mm_sub_pd(iz0,jz0);
1344             dx11             = _mm_sub_pd(ix1,jx1);
1345             dy11             = _mm_sub_pd(iy1,jy1);
1346             dz11             = _mm_sub_pd(iz1,jz1);
1347             dx12             = _mm_sub_pd(ix1,jx2);
1348             dy12             = _mm_sub_pd(iy1,jy2);
1349             dz12             = _mm_sub_pd(iz1,jz2);
1350             dx13             = _mm_sub_pd(ix1,jx3);
1351             dy13             = _mm_sub_pd(iy1,jy3);
1352             dz13             = _mm_sub_pd(iz1,jz3);
1353             dx21             = _mm_sub_pd(ix2,jx1);
1354             dy21             = _mm_sub_pd(iy2,jy1);
1355             dz21             = _mm_sub_pd(iz2,jz1);
1356             dx22             = _mm_sub_pd(ix2,jx2);
1357             dy22             = _mm_sub_pd(iy2,jy2);
1358             dz22             = _mm_sub_pd(iz2,jz2);
1359             dx23             = _mm_sub_pd(ix2,jx3);
1360             dy23             = _mm_sub_pd(iy2,jy3);
1361             dz23             = _mm_sub_pd(iz2,jz3);
1362             dx31             = _mm_sub_pd(ix3,jx1);
1363             dy31             = _mm_sub_pd(iy3,jy1);
1364             dz31             = _mm_sub_pd(iz3,jz1);
1365             dx32             = _mm_sub_pd(ix3,jx2);
1366             dy32             = _mm_sub_pd(iy3,jy2);
1367             dz32             = _mm_sub_pd(iz3,jz2);
1368             dx33             = _mm_sub_pd(ix3,jx3);
1369             dy33             = _mm_sub_pd(iy3,jy3);
1370             dz33             = _mm_sub_pd(iz3,jz3);
1371
1372             /* Calculate squared distance and things based on it */
1373             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1374             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1375             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1376             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1377             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1378             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1379             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1380             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1381             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1382             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1383
1384             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1385             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1386             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1387             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1388             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1389             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1390             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1391             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1392             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1393
1394             rinvsq00         = gmx_mm_inv_pd(rsq00);
1395             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1396             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1397             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1398             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1399             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1400             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1401             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1402             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1403             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1404
1405             fjx0             = _mm_setzero_pd();
1406             fjy0             = _mm_setzero_pd();
1407             fjz0             = _mm_setzero_pd();
1408             fjx1             = _mm_setzero_pd();
1409             fjy1             = _mm_setzero_pd();
1410             fjz1             = _mm_setzero_pd();
1411             fjx2             = _mm_setzero_pd();
1412             fjy2             = _mm_setzero_pd();
1413             fjz2             = _mm_setzero_pd();
1414             fjx3             = _mm_setzero_pd();
1415             fjy3             = _mm_setzero_pd();
1416             fjz3             = _mm_setzero_pd();
1417
1418             /**************************
1419              * CALCULATE INTERACTIONS *
1420              **************************/
1421
1422             /* LENNARD-JONES DISPERSION/REPULSION */
1423
1424             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1425             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1426
1427             fscal            = fvdw;
1428
1429             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1430
1431             /* Update vectorial force */
1432             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1433             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1434             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1435
1436             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1437             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1438             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1439
1440             /**************************
1441              * CALCULATE INTERACTIONS *
1442              **************************/
1443
1444             /* REACTION-FIELD ELECTROSTATICS */
1445             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1446
1447             fscal            = felec;
1448
1449             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1450
1451             /* Update vectorial force */
1452             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1453             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1454             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1455
1456             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1457             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1458             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1459
1460             /**************************
1461              * CALCULATE INTERACTIONS *
1462              **************************/
1463
1464             /* REACTION-FIELD ELECTROSTATICS */
1465             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1466
1467             fscal            = felec;
1468
1469             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1470
1471             /* Update vectorial force */
1472             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1473             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1474             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1475
1476             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1477             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1478             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1479
1480             /**************************
1481              * CALCULATE INTERACTIONS *
1482              **************************/
1483
1484             /* REACTION-FIELD ELECTROSTATICS */
1485             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1486
1487             fscal            = felec;
1488
1489             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1490
1491             /* Update vectorial force */
1492             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1493             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1494             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1495
1496             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1497             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1498             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1499
1500             /**************************
1501              * CALCULATE INTERACTIONS *
1502              **************************/
1503
1504             /* REACTION-FIELD ELECTROSTATICS */
1505             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1506
1507             fscal            = felec;
1508
1509             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1510
1511             /* Update vectorial force */
1512             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1513             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1514             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1515
1516             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1517             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1518             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1519
1520             /**************************
1521              * CALCULATE INTERACTIONS *
1522              **************************/
1523
1524             /* REACTION-FIELD ELECTROSTATICS */
1525             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1526
1527             fscal            = felec;
1528
1529             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1530
1531             /* Update vectorial force */
1532             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1533             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1534             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1535
1536             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1537             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1538             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1539
1540             /**************************
1541              * CALCULATE INTERACTIONS *
1542              **************************/
1543
1544             /* REACTION-FIELD ELECTROSTATICS */
1545             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1546
1547             fscal            = felec;
1548
1549             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1550
1551             /* Update vectorial force */
1552             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1553             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1554             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1555
1556             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1557             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1558             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1559
1560             /**************************
1561              * CALCULATE INTERACTIONS *
1562              **************************/
1563
1564             /* REACTION-FIELD ELECTROSTATICS */
1565             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1566
1567             fscal            = felec;
1568
1569             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1570
1571             /* Update vectorial force */
1572             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1573             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1574             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1575
1576             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1577             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1578             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1579
1580             /**************************
1581              * CALCULATE INTERACTIONS *
1582              **************************/
1583
1584             /* REACTION-FIELD ELECTROSTATICS */
1585             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1586
1587             fscal            = felec;
1588
1589             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1590
1591             /* Update vectorial force */
1592             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1593             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1594             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1595
1596             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1597             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1598             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1599
1600             /**************************
1601              * CALCULATE INTERACTIONS *
1602              **************************/
1603
1604             /* REACTION-FIELD ELECTROSTATICS */
1605             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1606
1607             fscal            = felec;
1608
1609             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1610
1611             /* Update vectorial force */
1612             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1613             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1614             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1615
1616             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1617             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1618             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1619
1620             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1621
1622             /* Inner loop uses 303 flops */
1623         }
1624
1625         /* End of innermost loop */
1626
1627         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1628                                               f+i_coord_offset,fshift+i_shift_offset);
1629
1630         /* Increment number of inner iterations */
1631         inneriter                  += j_index_end - j_index_start;
1632
1633         /* Outer loop uses 24 flops */
1634     }
1635
1636     /* Increment number of outer iterations */
1637     outeriter        += nri;
1638
1639     /* Update outer/inner flops */
1640
1641     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*303);
1642 }