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 "config.h"
  39
  40 #include <math.h>
  41
  42 #include "../nb_kernel.h"
  43 #include "types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "nrnb.h"
  46
  47 #include "gromacs/simd/math_x86_avx_128_fma_double.h"
  48 #include "kernelutil_x86_avx_128_fma_double.h"
  49
  50 /*
  51  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW4W4_VF_avx_128_fma_double
  52  * Electrostatics interaction: ReactionField
  53  * VdW interaction:            LennardJones
  54  * Geometry:                   Water4-Water4
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecRF_VdwLJ_GeomW4W4_VF_avx_128_fma_double
  59                     (t_nblist                    * gmx_restrict       nlist,
  60                      rvec                        * gmx_restrict          xx,
  61                      rvec                        * gmx_restrict          ff,
  62                      t_forcerec                  * gmx_restrict          fr,
  63                      t_mdatoms                   * gmx_restrict     mdatoms,
  64                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  65                      t_nrnb                      * gmx_restrict        nrnb)
  66 {
  67     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  68      * just 0 for non-waters.
  69      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  70      * jnr indices corresponding to data put in the four positions in the SIMD register.
  71      */
  72     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  73     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  74     int              jnrA,jnrB;
  75     int              j_coord_offsetA,j_coord_offsetB;
  76     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  77     real             rcutoff_scalar;
  78     real             *shiftvec,*fshift,*x,*f;
  79     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  80     int              vdwioffset0;
  81     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  82     int              vdwioffset1;
  83     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  84     int              vdwioffset2;
  85     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  86     int              vdwioffset3;
  87     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
  88     int              vdwjidx0A,vdwjidx0B;
  89     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  90     int              vdwjidx1A,vdwjidx1B;
  91     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  92     int              vdwjidx2A,vdwjidx2B;
  93     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  94     int              vdwjidx3A,vdwjidx3B;
  95     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
  96     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  97     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  98     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  99     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 100     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 101     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 102     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 103     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 104     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 105     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 106     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 107     real             *charge;
 108     int              nvdwtype;
 109     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 110     int              *vdwtype;
 111     real             *vdwparam;
 112     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 113     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 114     __m128d          dummy_mask,cutoff_mask;
 115     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 116     __m128d          one     = _mm_set1_pd(1.0);
 117     __m128d          two     = _mm_set1_pd(2.0);
 118     x                = xx[0];
 119     f                = ff[0];
 120
 121     nri              = nlist->nri;
 122     iinr             = nlist->iinr;
 123     jindex           = nlist->jindex;
 124     jjnr             = nlist->jjnr;
 125     shiftidx         = nlist->shift;
 126     gid              = nlist->gid;
 127     shiftvec         = fr->shift_vec[0];
 128     fshift           = fr->fshift[0];
 129     facel            = _mm_set1_pd(fr->epsfac);
 130     charge           = mdatoms->chargeA;
 131     krf              = _mm_set1_pd(fr->ic->k_rf);
 132     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 133     crf              = _mm_set1_pd(fr->ic->c_rf);
 134     nvdwtype         = fr->ntype;
 135     vdwparam         = fr->nbfp;
 136     vdwtype          = mdatoms->typeA;
 137
 138     /* Setup water-specific parameters */
 139     inr              = nlist->iinr[0];
 140     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 141     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 142     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 143     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 144
 145     jq1              = _mm_set1_pd(charge[inr+1]);
 146     jq2              = _mm_set1_pd(charge[inr+2]);
 147     jq3              = _mm_set1_pd(charge[inr+3]);
 148     vdwjidx0A        = 2*vdwtype[inr+0];
 149     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 150     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 151     qq11             = _mm_mul_pd(iq1,jq1);
 152     qq12             = _mm_mul_pd(iq1,jq2);
 153     qq13             = _mm_mul_pd(iq1,jq3);
 154     qq21             = _mm_mul_pd(iq2,jq1);
 155     qq22             = _mm_mul_pd(iq2,jq2);
 156     qq23             = _mm_mul_pd(iq2,jq3);
 157     qq31             = _mm_mul_pd(iq3,jq1);
 158     qq32             = _mm_mul_pd(iq3,jq2);
 159     qq33             = _mm_mul_pd(iq3,jq3);
 160
 161     /* Avoid stupid compiler warnings */
 162     jnrA = jnrB = 0;
 163     j_coord_offsetA = 0;
 164     j_coord_offsetB = 0;
 165
 166     outeriter        = 0;
 167     inneriter        = 0;
 168
 169     /* Start outer loop over neighborlists */
 170     for(iidx=0; iidx<nri; iidx++)
 171     {
 172         /* Load shift vector for this list */
 173         i_shift_offset   = DIM*shiftidx[iidx];
 174
 175         /* Load limits for loop over neighbors */
 176         j_index_start    = jindex[iidx];
 177         j_index_end      = jindex[iidx+1];
 178
 179         /* Get outer coordinate index */
 180         inr              = iinr[iidx];
 181         i_coord_offset   = DIM*inr;
 182
 183         /* Load i particle coords and add shift vector */
 184         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 185                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
 186
 187         fix0             = _mm_setzero_pd();
 188         fiy0             = _mm_setzero_pd();
 189         fiz0             = _mm_setzero_pd();
 190         fix1             = _mm_setzero_pd();
 191         fiy1             = _mm_setzero_pd();
 192         fiz1             = _mm_setzero_pd();
 193         fix2             = _mm_setzero_pd();
 194         fiy2             = _mm_setzero_pd();
 195         fiz2             = _mm_setzero_pd();
 196         fix3             = _mm_setzero_pd();
 197         fiy3             = _mm_setzero_pd();
 198         fiz3             = _mm_setzero_pd();
 199
 200         /* Reset potential sums */
 201         velecsum         = _mm_setzero_pd();
 202         vvdwsum          = _mm_setzero_pd();
 203
 204         /* Start inner kernel loop */
 205         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 206         {
 207
 208             /* Get j neighbor index, and coordinate index */
 209             jnrA             = jjnr[jidx];
 210             jnrB             = jjnr[jidx+1];
 211             j_coord_offsetA  = DIM*jnrA;
 212             j_coord_offsetB  = DIM*jnrB;
 213
 214             /* load j atom coordinates */
 215             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 216                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 217                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 218
 219             /* Calculate displacement vector */
 220             dx00             = _mm_sub_pd(ix0,jx0);
 221             dy00             = _mm_sub_pd(iy0,jy0);
 222             dz00             = _mm_sub_pd(iz0,jz0);
 223             dx11             = _mm_sub_pd(ix1,jx1);
 224             dy11             = _mm_sub_pd(iy1,jy1);
 225             dz11             = _mm_sub_pd(iz1,jz1);
 226             dx12             = _mm_sub_pd(ix1,jx2);
 227             dy12             = _mm_sub_pd(iy1,jy2);
 228             dz12             = _mm_sub_pd(iz1,jz2);
 229             dx13             = _mm_sub_pd(ix1,jx3);
 230             dy13             = _mm_sub_pd(iy1,jy3);
 231             dz13             = _mm_sub_pd(iz1,jz3);
 232             dx21             = _mm_sub_pd(ix2,jx1);
 233             dy21             = _mm_sub_pd(iy2,jy1);
 234             dz21             = _mm_sub_pd(iz2,jz1);
 235             dx22             = _mm_sub_pd(ix2,jx2);
 236             dy22             = _mm_sub_pd(iy2,jy2);
 237             dz22             = _mm_sub_pd(iz2,jz2);
 238             dx23             = _mm_sub_pd(ix2,jx3);
 239             dy23             = _mm_sub_pd(iy2,jy3);
 240             dz23             = _mm_sub_pd(iz2,jz3);
 241             dx31             = _mm_sub_pd(ix3,jx1);
 242             dy31             = _mm_sub_pd(iy3,jy1);
 243             dz31             = _mm_sub_pd(iz3,jz1);
 244             dx32             = _mm_sub_pd(ix3,jx2);
 245             dy32             = _mm_sub_pd(iy3,jy2);
 246             dz32             = _mm_sub_pd(iz3,jz2);
 247             dx33             = _mm_sub_pd(ix3,jx3);
 248             dy33             = _mm_sub_pd(iy3,jy3);
 249             dz33             = _mm_sub_pd(iz3,jz3);
 250
 251             /* Calculate squared distance and things based on it */
 252             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 253             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 254             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 255             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 256             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 257             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 258             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 259             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 260             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 261             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 262
 263             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 264             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 265             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 266             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 267             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 268             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 269             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 270             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 271             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 272
 273             rinvsq00         = gmx_mm_inv_pd(rsq00);
 274             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 275             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 276             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 277             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 278             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 279             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 280             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 281             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 282             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 283
 284             fjx0             = _mm_setzero_pd();
 285             fjy0             = _mm_setzero_pd();
 286             fjz0             = _mm_setzero_pd();
 287             fjx1             = _mm_setzero_pd();
 288             fjy1             = _mm_setzero_pd();
 289             fjz1             = _mm_setzero_pd();
 290             fjx2             = _mm_setzero_pd();
 291             fjy2             = _mm_setzero_pd();
 292             fjz2             = _mm_setzero_pd();
 293             fjx3             = _mm_setzero_pd();
 294             fjy3             = _mm_setzero_pd();
 295             fjz3             = _mm_setzero_pd();
 296
 297             /**************************
 298              * CALCULATE INTERACTIONS *
 299              **************************/
 300
 301             /* LENNARD-JONES DISPERSION/REPULSION */
 302
 303             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 304             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 305             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 306             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 307             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 308
 309             /* Update potential sum for this i atom from the interaction with this j atom. */
 310             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 311
 312             fscal            = fvdw;
 313
 314             /* Update vectorial force */
 315             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 316             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 317             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 318
 319             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 320             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 321             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 322
 323             /**************************
 324              * CALCULATE INTERACTIONS *
 325              **************************/
 326
 327             /* REACTION-FIELD ELECTROSTATICS */
 328             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 329             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 330
 331             /* Update potential sum for this i atom from the interaction with this j atom. */
 332             velecsum         = _mm_add_pd(velecsum,velec);
 333
 334             fscal            = felec;
 335
 336             /* Update vectorial force */
 337             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 338             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 339             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 340
 341             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 342             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 343             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 344
 345             /**************************
 346              * CALCULATE INTERACTIONS *
 347              **************************/
 348
 349             /* REACTION-FIELD ELECTROSTATICS */
 350             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 351             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 352
 353             /* Update potential sum for this i atom from the interaction with this j atom. */
 354             velecsum         = _mm_add_pd(velecsum,velec);
 355
 356             fscal            = felec;
 357
 358             /* Update vectorial force */
 359             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 360             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 361             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 362
 363             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 364             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 365             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 366
 367             /**************************
 368              * CALCULATE INTERACTIONS *
 369              **************************/
 370
 371             /* REACTION-FIELD ELECTROSTATICS */
 372             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 373             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 374
 375             /* Update potential sum for this i atom from the interaction with this j atom. */
 376             velecsum         = _mm_add_pd(velecsum,velec);
 377
 378             fscal            = felec;
 379
 380             /* Update vectorial force */
 381             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 382             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 383             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 384
 385             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 386             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 387             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 388
 389             /**************************
 390              * CALCULATE INTERACTIONS *
 391              **************************/
 392
 393             /* REACTION-FIELD ELECTROSTATICS */
 394             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 395             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 396
 397             /* Update potential sum for this i atom from the interaction with this j atom. */
 398             velecsum         = _mm_add_pd(velecsum,velec);
 399
 400             fscal            = felec;
 401
 402             /* Update vectorial force */
 403             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 404             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 405             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 406
 407             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 408             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 409             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 410
 411             /**************************
 412              * CALCULATE INTERACTIONS *
 413              **************************/
 414
 415             /* REACTION-FIELD ELECTROSTATICS */
 416             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 417             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 418
 419             /* Update potential sum for this i atom from the interaction with this j atom. */
 420             velecsum         = _mm_add_pd(velecsum,velec);
 421
 422             fscal            = felec;
 423
 424             /* Update vectorial force */
 425             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 426             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 427             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 428
 429             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 430             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 431             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 432
 433             /**************************
 434              * CALCULATE INTERACTIONS *
 435              **************************/
 436
 437             /* REACTION-FIELD ELECTROSTATICS */
 438             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 439             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 440
 441             /* Update potential sum for this i atom from the interaction with this j atom. */
 442             velecsum         = _mm_add_pd(velecsum,velec);
 443
 444             fscal            = felec;
 445
 446             /* Update vectorial force */
 447             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 448             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 449             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 450
 451             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 452             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 453             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 454
 455             /**************************
 456              * CALCULATE INTERACTIONS *
 457              **************************/
 458
 459             /* REACTION-FIELD ELECTROSTATICS */
 460             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
 461             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
 462
 463             /* Update potential sum for this i atom from the interaction with this j atom. */
 464             velecsum         = _mm_add_pd(velecsum,velec);
 465
 466             fscal            = felec;
 467
 468             /* Update vectorial force */
 469             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 470             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 471             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 472
 473             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 474             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 475             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 476
 477             /**************************
 478              * CALCULATE INTERACTIONS *
 479              **************************/
 480
 481             /* REACTION-FIELD ELECTROSTATICS */
 482             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
 483             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
 484
 485             /* Update potential sum for this i atom from the interaction with this j atom. */
 486             velecsum         = _mm_add_pd(velecsum,velec);
 487
 488             fscal            = felec;
 489
 490             /* Update vectorial force */
 491             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 492             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 493             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 494
 495             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 496             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 497             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 498
 499             /**************************
 500              * CALCULATE INTERACTIONS *
 501              **************************/
 502
 503             /* REACTION-FIELD ELECTROSTATICS */
 504             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
 505             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
 506
 507             /* Update potential sum for this i atom from the interaction with this j atom. */
 508             velecsum         = _mm_add_pd(velecsum,velec);
 509
 510             fscal            = felec;
 511
 512             /* Update vectorial force */
 513             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 514             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 515             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 516
 517             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 518             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 519             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 520
 521             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);
 522
 523             /* Inner loop uses 353 flops */
 524         }
 525
 526         if(jidx<j_index_end)
 527         {
 528
 529             jnrA             = jjnr[jidx];
 530             j_coord_offsetA  = DIM*jnrA;
 531
 532             /* load j atom coordinates */
 533             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 534                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
 535                                               &jy2,&jz2,&jx3,&jy3,&jz3);
 536
 537             /* Calculate displacement vector */
 538             dx00             = _mm_sub_pd(ix0,jx0);
 539             dy00             = _mm_sub_pd(iy0,jy0);
 540             dz00             = _mm_sub_pd(iz0,jz0);
 541             dx11             = _mm_sub_pd(ix1,jx1);
 542             dy11             = _mm_sub_pd(iy1,jy1);
 543             dz11             = _mm_sub_pd(iz1,jz1);
 544             dx12             = _mm_sub_pd(ix1,jx2);
 545             dy12             = _mm_sub_pd(iy1,jy2);
 546             dz12             = _mm_sub_pd(iz1,jz2);
 547             dx13             = _mm_sub_pd(ix1,jx3);
 548             dy13             = _mm_sub_pd(iy1,jy3);
 549             dz13             = _mm_sub_pd(iz1,jz3);
 550             dx21             = _mm_sub_pd(ix2,jx1);
 551             dy21             = _mm_sub_pd(iy2,jy1);
 552             dz21             = _mm_sub_pd(iz2,jz1);
 553             dx22             = _mm_sub_pd(ix2,jx2);
 554             dy22             = _mm_sub_pd(iy2,jy2);
 555             dz22             = _mm_sub_pd(iz2,jz2);
 556             dx23             = _mm_sub_pd(ix2,jx3);
 557             dy23             = _mm_sub_pd(iy2,jy3);
 558             dz23             = _mm_sub_pd(iz2,jz3);
 559             dx31             = _mm_sub_pd(ix3,jx1);
 560             dy31             = _mm_sub_pd(iy3,jy1);
 561             dz31             = _mm_sub_pd(iz3,jz1);
 562             dx32             = _mm_sub_pd(ix3,jx2);
 563             dy32             = _mm_sub_pd(iy3,jy2);
 564             dz32             = _mm_sub_pd(iz3,jz2);
 565             dx33             = _mm_sub_pd(ix3,jx3);
 566             dy33             = _mm_sub_pd(iy3,jy3);
 567             dz33             = _mm_sub_pd(iz3,jz3);
 568
 569             /* Calculate squared distance and things based on it */
 570             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 571             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 572             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 573             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
 574             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 575             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 576             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
 577             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
 578             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
 579             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
 580
 581             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 582             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 583             rinv13           = gmx_mm_invsqrt_pd(rsq13);
 584             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 585             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 586             rinv23           = gmx_mm_invsqrt_pd(rsq23);
 587             rinv31           = gmx_mm_invsqrt_pd(rsq31);
 588             rinv32           = gmx_mm_invsqrt_pd(rsq32);
 589             rinv33           = gmx_mm_invsqrt_pd(rsq33);
 590
 591             rinvsq00         = gmx_mm_inv_pd(rsq00);
 592             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 593             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 594             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
 595             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 596             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 597             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
 598             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
 599             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
 600             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
 601
 602             fjx0             = _mm_setzero_pd();
 603             fjy0             = _mm_setzero_pd();
 604             fjz0             = _mm_setzero_pd();
 605             fjx1             = _mm_setzero_pd();
 606             fjy1             = _mm_setzero_pd();
 607             fjz1             = _mm_setzero_pd();
 608             fjx2             = _mm_setzero_pd();
 609             fjy2             = _mm_setzero_pd();
 610             fjz2             = _mm_setzero_pd();
 611             fjx3             = _mm_setzero_pd();
 612             fjy3             = _mm_setzero_pd();
 613             fjz3             = _mm_setzero_pd();
 614
 615             /**************************
 616              * CALCULATE INTERACTIONS *
 617              **************************/
 618
 619             /* LENNARD-JONES DISPERSION/REPULSION */
 620
 621             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 622             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 623             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 624             vvdw             = _mm_msub_pd( vvdw12,one_twelfth, _mm_mul_pd(vvdw6,one_sixth) );
 625             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 626
 627             /* Update potential sum for this i atom from the interaction with this j atom. */
 628             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 629             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 630
 631             fscal            = fvdw;
 632
 633             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 634
 635             /* Update vectorial force */
 636             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 637             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 638             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 639
 640             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 641             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 642             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 643
 644             /**************************
 645              * CALCULATE INTERACTIONS *
 646              **************************/
 647
 648             /* REACTION-FIELD ELECTROSTATICS */
 649             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 650             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 651
 652             /* Update potential sum for this i atom from the interaction with this j atom. */
 653             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 654             velecsum         = _mm_add_pd(velecsum,velec);
 655
 656             fscal            = felec;
 657
 658             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 659
 660             /* Update vectorial force */
 661             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 662             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 663             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 664
 665             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 666             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 667             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 668
 669             /**************************
 670              * CALCULATE INTERACTIONS *
 671              **************************/
 672
 673             /* REACTION-FIELD ELECTROSTATICS */
 674             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 675             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 676
 677             /* Update potential sum for this i atom from the interaction with this j atom. */
 678             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 679             velecsum         = _mm_add_pd(velecsum,velec);
 680
 681             fscal            = felec;
 682
 683             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 684
 685             /* Update vectorial force */
 686             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 687             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 688             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 689
 690             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 691             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 692             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 693
 694             /**************************
 695              * CALCULATE INTERACTIONS *
 696              **************************/
 697
 698             /* REACTION-FIELD ELECTROSTATICS */
 699             velec            = _mm_mul_pd(qq13,_mm_sub_pd(_mm_macc_pd(krf,rsq13,rinv13),crf));
 700             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
 701
 702             /* Update potential sum for this i atom from the interaction with this j atom. */
 703             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 704             velecsum         = _mm_add_pd(velecsum,velec);
 705
 706             fscal            = felec;
 707
 708             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 709
 710             /* Update vectorial force */
 711             fix1             = _mm_macc_pd(dx13,fscal,fix1);
 712             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
 713             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
 714
 715             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
 716             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
 717             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
 718
 719             /**************************
 720              * CALCULATE INTERACTIONS *
 721              **************************/
 722
 723             /* REACTION-FIELD ELECTROSTATICS */
 724             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 725             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 726
 727             /* Update potential sum for this i atom from the interaction with this j atom. */
 728             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 729             velecsum         = _mm_add_pd(velecsum,velec);
 730
 731             fscal            = felec;
 732
 733             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 734
 735             /* Update vectorial force */
 736             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 737             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 738             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 739
 740             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 741             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 742             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 743
 744             /**************************
 745              * CALCULATE INTERACTIONS *
 746              **************************/
 747
 748             /* REACTION-FIELD ELECTROSTATICS */
 749             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 750             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 751
 752             /* Update potential sum for this i atom from the interaction with this j atom. */
 753             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 754             velecsum         = _mm_add_pd(velecsum,velec);
 755
 756             fscal            = felec;
 757
 758             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 759
 760             /* Update vectorial force */
 761             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 762             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 763             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 764
 765             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 766             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 767             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 768
 769             /**************************
 770              * CALCULATE INTERACTIONS *
 771              **************************/
 772
 773             /* REACTION-FIELD ELECTROSTATICS */
 774             velec            = _mm_mul_pd(qq23,_mm_sub_pd(_mm_macc_pd(krf,rsq23,rinv23),crf));
 775             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
 776
 777             /* Update potential sum for this i atom from the interaction with this j atom. */
 778             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 779             velecsum         = _mm_add_pd(velecsum,velec);
 780
 781             fscal            = felec;
 782
 783             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 784
 785             /* Update vectorial force */
 786             fix2             = _mm_macc_pd(dx23,fscal,fix2);
 787             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
 788             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
 789
 790             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
 791             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
 792             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
 793
 794             /**************************
 795              * CALCULATE INTERACTIONS *
 796              **************************/
 797
 798             /* REACTION-FIELD ELECTROSTATICS */
 799             velec            = _mm_mul_pd(qq31,_mm_sub_pd(_mm_macc_pd(krf,rsq31,rinv31),crf));
 800             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
 801
 802             /* Update potential sum for this i atom from the interaction with this j atom. */
 803             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 804             velecsum         = _mm_add_pd(velecsum,velec);
 805
 806             fscal            = felec;
 807
 808             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 809
 810             /* Update vectorial force */
 811             fix3             = _mm_macc_pd(dx31,fscal,fix3);
 812             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
 813             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
 814
 815             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
 816             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
 817             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
 818
 819             /**************************
 820              * CALCULATE INTERACTIONS *
 821              **************************/
 822
 823             /* REACTION-FIELD ELECTROSTATICS */
 824             velec            = _mm_mul_pd(qq32,_mm_sub_pd(_mm_macc_pd(krf,rsq32,rinv32),crf));
 825             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
 826
 827             /* Update potential sum for this i atom from the interaction with this j atom. */
 828             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 829             velecsum         = _mm_add_pd(velecsum,velec);
 830
 831             fscal            = felec;
 832
 833             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 834
 835             /* Update vectorial force */
 836             fix3             = _mm_macc_pd(dx32,fscal,fix3);
 837             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
 838             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
 839
 840             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
 841             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
 842             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
 843
 844             /**************************
 845              * CALCULATE INTERACTIONS *
 846              **************************/
 847
 848             /* REACTION-FIELD ELECTROSTATICS */
 849             velec            = _mm_mul_pd(qq33,_mm_sub_pd(_mm_macc_pd(krf,rsq33,rinv33),crf));
 850             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
 851
 852             /* Update potential sum for this i atom from the interaction with this j atom. */
 853             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 854             velecsum         = _mm_add_pd(velecsum,velec);
 855
 856             fscal            = felec;
 857
 858             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 859
 860             /* Update vectorial force */
 861             fix3             = _mm_macc_pd(dx33,fscal,fix3);
 862             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
 863             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
 864
 865             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
 866             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
 867             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
 868
 869             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
 870
 871             /* Inner loop uses 353 flops */
 872         }
 873
 874         /* End of innermost loop */
 875
 876         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
 877                                               f+i_coord_offset,fshift+i_shift_offset);
 878
 879         ggid                        = gid[iidx];
 880         /* Update potential energies */
 881         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
 882         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
 883
 884         /* Increment number of inner iterations */
 885         inneriter                  += j_index_end - j_index_start;
 886
 887         /* Outer loop uses 26 flops */
 888     }
 889
 890     /* Increment number of outer iterations */
 891     outeriter        += nri;
 892
 893     /* Update outer/inner flops */
 894
 895     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_VF,outeriter*26 + inneriter*353);
 896 }
 897 /*
 898  * Gromacs nonbonded kernel:   nb_kernel_ElecRF_VdwLJ_GeomW4W4_F_avx_128_fma_double
 899  * Electrostatics interaction: ReactionField
 900  * VdW interaction:            LennardJones
 901  * Geometry:                   Water4-Water4
 902  * Calculate force/pot:        Force
 903  */
 904 void
 905 nb_kernel_ElecRF_VdwLJ_GeomW4W4_F_avx_128_fma_double
 906                     (t_nblist                    * gmx_restrict       nlist,
 907                      rvec                        * gmx_restrict          xx,
 908                      rvec                        * gmx_restrict          ff,
 909                      t_forcerec                  * gmx_restrict          fr,
 910                      t_mdatoms                   * gmx_restrict     mdatoms,
 911                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
 912                      t_nrnb                      * gmx_restrict        nrnb)
 913 {
 914     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
 915      * just 0 for non-waters.
 916      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
 917      * jnr indices corresponding to data put in the four positions in the SIMD register.
 918      */
 919     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
 920     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
 921     int              jnrA,jnrB;
 922     int              j_coord_offsetA,j_coord_offsetB;
 923     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
 924     real             rcutoff_scalar;
 925     real             *shiftvec,*fshift,*x,*f;
 926     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
 927     int              vdwioffset0;
 928     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
 929     int              vdwioffset1;
 930     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
 931     int              vdwioffset2;
 932     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
 933     int              vdwioffset3;
 934     __m128d          ix3,iy3,iz3,fix3,fiy3,fiz3,iq3,isai3;
 935     int              vdwjidx0A,vdwjidx0B;
 936     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
 937     int              vdwjidx1A,vdwjidx1B;
 938     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
 939     int              vdwjidx2A,vdwjidx2B;
 940     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
 941     int              vdwjidx3A,vdwjidx3B;
 942     __m128d          jx3,jy3,jz3,fjx3,fjy3,fjz3,jq3,isaj3;
 943     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
 944     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
 945     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
 946     __m128d          dx13,dy13,dz13,rsq13,rinv13,rinvsq13,r13,qq13,c6_13,c12_13;
 947     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 948     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 949     __m128d          dx23,dy23,dz23,rsq23,rinv23,rinvsq23,r23,qq23,c6_23,c12_23;
 950     __m128d          dx31,dy31,dz31,rsq31,rinv31,rinvsq31,r31,qq31,c6_31,c12_31;
 951     __m128d          dx32,dy32,dz32,rsq32,rinv32,rinvsq32,r32,qq32,c6_32,c12_32;
 952     __m128d          dx33,dy33,dz33,rsq33,rinv33,rinvsq33,r33,qq33,c6_33,c12_33;
 953     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 954     real             *charge;
 955     int              nvdwtype;
 956     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 957     int              *vdwtype;
 958     real             *vdwparam;
 959     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 960     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 961     __m128d          dummy_mask,cutoff_mask;
 962     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 963     __m128d          one     = _mm_set1_pd(1.0);
 964     __m128d          two     = _mm_set1_pd(2.0);
 965     x                = xx[0];
 966     f                = ff[0];
 967
 968     nri              = nlist->nri;
 969     iinr             = nlist->iinr;
 970     jindex           = nlist->jindex;
 971     jjnr             = nlist->jjnr;
 972     shiftidx         = nlist->shift;
 973     gid              = nlist->gid;
 974     shiftvec         = fr->shift_vec[0];
 975     fshift           = fr->fshift[0];
 976     facel            = _mm_set1_pd(fr->epsfac);
 977     charge           = mdatoms->chargeA;
 978     krf              = _mm_set1_pd(fr->ic->k_rf);
 979     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 980     crf              = _mm_set1_pd(fr->ic->c_rf);
 981     nvdwtype         = fr->ntype;
 982     vdwparam         = fr->nbfp;
 983     vdwtype          = mdatoms->typeA;
 984
 985     /* Setup water-specific parameters */
 986     inr              = nlist->iinr[0];
 987     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 988     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 989     iq3              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+3]));
 990     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 991
 992     jq1              = _mm_set1_pd(charge[inr+1]);
 993     jq2              = _mm_set1_pd(charge[inr+2]);
 994     jq3              = _mm_set1_pd(charge[inr+3]);
 995     vdwjidx0A        = 2*vdwtype[inr+0];
 996     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 997     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 998     qq11             = _mm_mul_pd(iq1,jq1);
 999     qq12             = _mm_mul_pd(iq1,jq2);
1000     qq13             = _mm_mul_pd(iq1,jq3);
1001     qq21             = _mm_mul_pd(iq2,jq1);
1002     qq22             = _mm_mul_pd(iq2,jq2);
1003     qq23             = _mm_mul_pd(iq2,jq3);
1004     qq31             = _mm_mul_pd(iq3,jq1);
1005     qq32             = _mm_mul_pd(iq3,jq2);
1006     qq33             = _mm_mul_pd(iq3,jq3);
1007
1008     /* Avoid stupid compiler warnings */
1009     jnrA = jnrB = 0;
1010     j_coord_offsetA = 0;
1011     j_coord_offsetB = 0;
1012
1013     outeriter        = 0;
1014     inneriter        = 0;
1015
1016     /* Start outer loop over neighborlists */
1017     for(iidx=0; iidx<nri; iidx++)
1018     {
1019         /* Load shift vector for this list */
1020         i_shift_offset   = DIM*shiftidx[iidx];
1021
1022         /* Load limits for loop over neighbors */
1023         j_index_start    = jindex[iidx];
1024         j_index_end      = jindex[iidx+1];
1025
1026         /* Get outer coordinate index */
1027         inr              = iinr[iidx];
1028         i_coord_offset   = DIM*inr;
1029
1030         /* Load i particle coords and add shift vector */
1031         gmx_mm_load_shift_and_4rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1032                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2,&ix3,&iy3,&iz3);
1033
1034         fix0             = _mm_setzero_pd();
1035         fiy0             = _mm_setzero_pd();
1036         fiz0             = _mm_setzero_pd();
1037         fix1             = _mm_setzero_pd();
1038         fiy1             = _mm_setzero_pd();
1039         fiz1             = _mm_setzero_pd();
1040         fix2             = _mm_setzero_pd();
1041         fiy2             = _mm_setzero_pd();
1042         fiz2             = _mm_setzero_pd();
1043         fix3             = _mm_setzero_pd();
1044         fiy3             = _mm_setzero_pd();
1045         fiz3             = _mm_setzero_pd();
1046
1047         /* Start inner kernel loop */
1048         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1049         {
1050
1051             /* Get j neighbor index, and coordinate index */
1052             jnrA             = jjnr[jidx];
1053             jnrB             = jjnr[jidx+1];
1054             j_coord_offsetA  = DIM*jnrA;
1055             j_coord_offsetB  = DIM*jnrB;
1056
1057             /* load j atom coordinates */
1058             gmx_mm_load_4rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1059                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1060                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1061
1062             /* Calculate displacement vector */
1063             dx00             = _mm_sub_pd(ix0,jx0);
1064             dy00             = _mm_sub_pd(iy0,jy0);
1065             dz00             = _mm_sub_pd(iz0,jz0);
1066             dx11             = _mm_sub_pd(ix1,jx1);
1067             dy11             = _mm_sub_pd(iy1,jy1);
1068             dz11             = _mm_sub_pd(iz1,jz1);
1069             dx12             = _mm_sub_pd(ix1,jx2);
1070             dy12             = _mm_sub_pd(iy1,jy2);
1071             dz12             = _mm_sub_pd(iz1,jz2);
1072             dx13             = _mm_sub_pd(ix1,jx3);
1073             dy13             = _mm_sub_pd(iy1,jy3);
1074             dz13             = _mm_sub_pd(iz1,jz3);
1075             dx21             = _mm_sub_pd(ix2,jx1);
1076             dy21             = _mm_sub_pd(iy2,jy1);
1077             dz21             = _mm_sub_pd(iz2,jz1);
1078             dx22             = _mm_sub_pd(ix2,jx2);
1079             dy22             = _mm_sub_pd(iy2,jy2);
1080             dz22             = _mm_sub_pd(iz2,jz2);
1081             dx23             = _mm_sub_pd(ix2,jx3);
1082             dy23             = _mm_sub_pd(iy2,jy3);
1083             dz23             = _mm_sub_pd(iz2,jz3);
1084             dx31             = _mm_sub_pd(ix3,jx1);
1085             dy31             = _mm_sub_pd(iy3,jy1);
1086             dz31             = _mm_sub_pd(iz3,jz1);
1087             dx32             = _mm_sub_pd(ix3,jx2);
1088             dy32             = _mm_sub_pd(iy3,jy2);
1089             dz32             = _mm_sub_pd(iz3,jz2);
1090             dx33             = _mm_sub_pd(ix3,jx3);
1091             dy33             = _mm_sub_pd(iy3,jy3);
1092             dz33             = _mm_sub_pd(iz3,jz3);
1093
1094             /* Calculate squared distance and things based on it */
1095             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1096             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1097             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1098             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1099             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1100             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1101             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1102             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1103             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1104             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1105
1106             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1107             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1108             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1109             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1110             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1111             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1112             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1113             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1114             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1115
1116             rinvsq00         = gmx_mm_inv_pd(rsq00);
1117             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1118             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1119             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1120             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1121             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1122             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1123             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1124             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1125             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1126
1127             fjx0             = _mm_setzero_pd();
1128             fjy0             = _mm_setzero_pd();
1129             fjz0             = _mm_setzero_pd();
1130             fjx1             = _mm_setzero_pd();
1131             fjy1             = _mm_setzero_pd();
1132             fjz1             = _mm_setzero_pd();
1133             fjx2             = _mm_setzero_pd();
1134             fjy2             = _mm_setzero_pd();
1135             fjz2             = _mm_setzero_pd();
1136             fjx3             = _mm_setzero_pd();
1137             fjy3             = _mm_setzero_pd();
1138             fjz3             = _mm_setzero_pd();
1139
1140             /**************************
1141              * CALCULATE INTERACTIONS *
1142              **************************/
1143
1144             /* LENNARD-JONES DISPERSION/REPULSION */
1145
1146             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1147             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1148
1149             fscal            = fvdw;
1150
1151             /* Update vectorial force */
1152             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1153             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1154             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1155
1156             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1157             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1158             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1159
1160             /**************************
1161              * CALCULATE INTERACTIONS *
1162              **************************/
1163
1164             /* REACTION-FIELD ELECTROSTATICS */
1165             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1166
1167             fscal            = felec;
1168
1169             /* Update vectorial force */
1170             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1171             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1172             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1173
1174             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1175             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1176             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1177
1178             /**************************
1179              * CALCULATE INTERACTIONS *
1180              **************************/
1181
1182             /* REACTION-FIELD ELECTROSTATICS */
1183             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1184
1185             fscal            = felec;
1186
1187             /* Update vectorial force */
1188             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1189             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1190             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1191
1192             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1193             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1194             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1195
1196             /**************************
1197              * CALCULATE INTERACTIONS *
1198              **************************/
1199
1200             /* REACTION-FIELD ELECTROSTATICS */
1201             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1202
1203             fscal            = felec;
1204
1205             /* Update vectorial force */
1206             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1207             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1208             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1209
1210             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1211             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1212             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1213
1214             /**************************
1215              * CALCULATE INTERACTIONS *
1216              **************************/
1217
1218             /* REACTION-FIELD ELECTROSTATICS */
1219             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1220
1221             fscal            = felec;
1222
1223             /* Update vectorial force */
1224             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1225             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1226             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1227
1228             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1229             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1230             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1231
1232             /**************************
1233              * CALCULATE INTERACTIONS *
1234              **************************/
1235
1236             /* REACTION-FIELD ELECTROSTATICS */
1237             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1238
1239             fscal            = felec;
1240
1241             /* Update vectorial force */
1242             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1243             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1244             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1245
1246             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1247             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1248             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1249
1250             /**************************
1251              * CALCULATE INTERACTIONS *
1252              **************************/
1253
1254             /* REACTION-FIELD ELECTROSTATICS */
1255             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1256
1257             fscal            = felec;
1258
1259             /* Update vectorial force */
1260             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1261             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1262             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1263
1264             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1265             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1266             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1267
1268             /**************************
1269              * CALCULATE INTERACTIONS *
1270              **************************/
1271
1272             /* REACTION-FIELD ELECTROSTATICS */
1273             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1274
1275             fscal            = felec;
1276
1277             /* Update vectorial force */
1278             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1279             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1280             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1281
1282             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1283             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1284             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1285
1286             /**************************
1287              * CALCULATE INTERACTIONS *
1288              **************************/
1289
1290             /* REACTION-FIELD ELECTROSTATICS */
1291             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1292
1293             fscal            = felec;
1294
1295             /* Update vectorial force */
1296             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1297             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1298             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1299
1300             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1301             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1302             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1303
1304             /**************************
1305              * CALCULATE INTERACTIONS *
1306              **************************/
1307
1308             /* REACTION-FIELD ELECTROSTATICS */
1309             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1310
1311             fscal            = felec;
1312
1313             /* Update vectorial force */
1314             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1315             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1316             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1317
1318             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1319             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1320             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1321
1322             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);
1323
1324             /* Inner loop uses 303 flops */
1325         }
1326
1327         if(jidx<j_index_end)
1328         {
1329
1330             jnrA             = jjnr[jidx];
1331             j_coord_offsetA  = DIM*jnrA;
1332
1333             /* load j atom coordinates */
1334             gmx_mm_load_4rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1335                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,
1336                                               &jy2,&jz2,&jx3,&jy3,&jz3);
1337
1338             /* Calculate displacement vector */
1339             dx00             = _mm_sub_pd(ix0,jx0);
1340             dy00             = _mm_sub_pd(iy0,jy0);
1341             dz00             = _mm_sub_pd(iz0,jz0);
1342             dx11             = _mm_sub_pd(ix1,jx1);
1343             dy11             = _mm_sub_pd(iy1,jy1);
1344             dz11             = _mm_sub_pd(iz1,jz1);
1345             dx12             = _mm_sub_pd(ix1,jx2);
1346             dy12             = _mm_sub_pd(iy1,jy2);
1347             dz12             = _mm_sub_pd(iz1,jz2);
1348             dx13             = _mm_sub_pd(ix1,jx3);
1349             dy13             = _mm_sub_pd(iy1,jy3);
1350             dz13             = _mm_sub_pd(iz1,jz3);
1351             dx21             = _mm_sub_pd(ix2,jx1);
1352             dy21             = _mm_sub_pd(iy2,jy1);
1353             dz21             = _mm_sub_pd(iz2,jz1);
1354             dx22             = _mm_sub_pd(ix2,jx2);
1355             dy22             = _mm_sub_pd(iy2,jy2);
1356             dz22             = _mm_sub_pd(iz2,jz2);
1357             dx23             = _mm_sub_pd(ix2,jx3);
1358             dy23             = _mm_sub_pd(iy2,jy3);
1359             dz23             = _mm_sub_pd(iz2,jz3);
1360             dx31             = _mm_sub_pd(ix3,jx1);
1361             dy31             = _mm_sub_pd(iy3,jy1);
1362             dz31             = _mm_sub_pd(iz3,jz1);
1363             dx32             = _mm_sub_pd(ix3,jx2);
1364             dy32             = _mm_sub_pd(iy3,jy2);
1365             dz32             = _mm_sub_pd(iz3,jz2);
1366             dx33             = _mm_sub_pd(ix3,jx3);
1367             dy33             = _mm_sub_pd(iy3,jy3);
1368             dz33             = _mm_sub_pd(iz3,jz3);
1369
1370             /* Calculate squared distance and things based on it */
1371             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1372             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1373             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1374             rsq13            = gmx_mm_calc_rsq_pd(dx13,dy13,dz13);
1375             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1376             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1377             rsq23            = gmx_mm_calc_rsq_pd(dx23,dy23,dz23);
1378             rsq31            = gmx_mm_calc_rsq_pd(dx31,dy31,dz31);
1379             rsq32            = gmx_mm_calc_rsq_pd(dx32,dy32,dz32);
1380             rsq33            = gmx_mm_calc_rsq_pd(dx33,dy33,dz33);
1381
1382             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1383             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1384             rinv13           = gmx_mm_invsqrt_pd(rsq13);
1385             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1386             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1387             rinv23           = gmx_mm_invsqrt_pd(rsq23);
1388             rinv31           = gmx_mm_invsqrt_pd(rsq31);
1389             rinv32           = gmx_mm_invsqrt_pd(rsq32);
1390             rinv33           = gmx_mm_invsqrt_pd(rsq33);
1391
1392             rinvsq00         = gmx_mm_inv_pd(rsq00);
1393             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1394             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1395             rinvsq13         = _mm_mul_pd(rinv13,rinv13);
1396             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1397             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1398             rinvsq23         = _mm_mul_pd(rinv23,rinv23);
1399             rinvsq31         = _mm_mul_pd(rinv31,rinv31);
1400             rinvsq32         = _mm_mul_pd(rinv32,rinv32);
1401             rinvsq33         = _mm_mul_pd(rinv33,rinv33);
1402
1403             fjx0             = _mm_setzero_pd();
1404             fjy0             = _mm_setzero_pd();
1405             fjz0             = _mm_setzero_pd();
1406             fjx1             = _mm_setzero_pd();
1407             fjy1             = _mm_setzero_pd();
1408             fjz1             = _mm_setzero_pd();
1409             fjx2             = _mm_setzero_pd();
1410             fjy2             = _mm_setzero_pd();
1411             fjz2             = _mm_setzero_pd();
1412             fjx3             = _mm_setzero_pd();
1413             fjy3             = _mm_setzero_pd();
1414             fjz3             = _mm_setzero_pd();
1415
1416             /**************************
1417              * CALCULATE INTERACTIONS *
1418              **************************/
1419
1420             /* LENNARD-JONES DISPERSION/REPULSION */
1421
1422             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1423             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1424
1425             fscal            = fvdw;
1426
1427             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1428
1429             /* Update vectorial force */
1430             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1431             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1432             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1433
1434             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1435             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1436             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1437
1438             /**************************
1439              * CALCULATE INTERACTIONS *
1440              **************************/
1441
1442             /* REACTION-FIELD ELECTROSTATICS */
1443             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1444
1445             fscal            = felec;
1446
1447             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1448
1449             /* Update vectorial force */
1450             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1451             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1452             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1453
1454             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1455             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1456             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1457
1458             /**************************
1459              * CALCULATE INTERACTIONS *
1460              **************************/
1461
1462             /* REACTION-FIELD ELECTROSTATICS */
1463             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1464
1465             fscal            = felec;
1466
1467             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1468
1469             /* Update vectorial force */
1470             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1471             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1472             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1473
1474             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1475             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1476             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1477
1478             /**************************
1479              * CALCULATE INTERACTIONS *
1480              **************************/
1481
1482             /* REACTION-FIELD ELECTROSTATICS */
1483             felec            = _mm_mul_pd(qq13,_mm_msub_pd(rinv13,rinvsq13,krf2));
1484
1485             fscal            = felec;
1486
1487             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1488
1489             /* Update vectorial force */
1490             fix1             = _mm_macc_pd(dx13,fscal,fix1);
1491             fiy1             = _mm_macc_pd(dy13,fscal,fiy1);
1492             fiz1             = _mm_macc_pd(dz13,fscal,fiz1);
1493
1494             fjx3             = _mm_macc_pd(dx13,fscal,fjx3);
1495             fjy3             = _mm_macc_pd(dy13,fscal,fjy3);
1496             fjz3             = _mm_macc_pd(dz13,fscal,fjz3);
1497
1498             /**************************
1499              * CALCULATE INTERACTIONS *
1500              **************************/
1501
1502             /* REACTION-FIELD ELECTROSTATICS */
1503             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1504
1505             fscal            = felec;
1506
1507             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1508
1509             /* Update vectorial force */
1510             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1511             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1512             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1513
1514             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1515             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1516             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1517
1518             /**************************
1519              * CALCULATE INTERACTIONS *
1520              **************************/
1521
1522             /* REACTION-FIELD ELECTROSTATICS */
1523             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1524
1525             fscal            = felec;
1526
1527             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1528
1529             /* Update vectorial force */
1530             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1531             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1532             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1533
1534             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1535             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1536             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1537
1538             /**************************
1539              * CALCULATE INTERACTIONS *
1540              **************************/
1541
1542             /* REACTION-FIELD ELECTROSTATICS */
1543             felec            = _mm_mul_pd(qq23,_mm_msub_pd(rinv23,rinvsq23,krf2));
1544
1545             fscal            = felec;
1546
1547             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1548
1549             /* Update vectorial force */
1550             fix2             = _mm_macc_pd(dx23,fscal,fix2);
1551             fiy2             = _mm_macc_pd(dy23,fscal,fiy2);
1552             fiz2             = _mm_macc_pd(dz23,fscal,fiz2);
1553
1554             fjx3             = _mm_macc_pd(dx23,fscal,fjx3);
1555             fjy3             = _mm_macc_pd(dy23,fscal,fjy3);
1556             fjz3             = _mm_macc_pd(dz23,fscal,fjz3);
1557
1558             /**************************
1559              * CALCULATE INTERACTIONS *
1560              **************************/
1561
1562             /* REACTION-FIELD ELECTROSTATICS */
1563             felec            = _mm_mul_pd(qq31,_mm_msub_pd(rinv31,rinvsq31,krf2));
1564
1565             fscal            = felec;
1566
1567             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1568
1569             /* Update vectorial force */
1570             fix3             = _mm_macc_pd(dx31,fscal,fix3);
1571             fiy3             = _mm_macc_pd(dy31,fscal,fiy3);
1572             fiz3             = _mm_macc_pd(dz31,fscal,fiz3);
1573
1574             fjx1             = _mm_macc_pd(dx31,fscal,fjx1);
1575             fjy1             = _mm_macc_pd(dy31,fscal,fjy1);
1576             fjz1             = _mm_macc_pd(dz31,fscal,fjz1);
1577
1578             /**************************
1579              * CALCULATE INTERACTIONS *
1580              **************************/
1581
1582             /* REACTION-FIELD ELECTROSTATICS */
1583             felec            = _mm_mul_pd(qq32,_mm_msub_pd(rinv32,rinvsq32,krf2));
1584
1585             fscal            = felec;
1586
1587             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1588
1589             /* Update vectorial force */
1590             fix3             = _mm_macc_pd(dx32,fscal,fix3);
1591             fiy3             = _mm_macc_pd(dy32,fscal,fiy3);
1592             fiz3             = _mm_macc_pd(dz32,fscal,fiz3);
1593
1594             fjx2             = _mm_macc_pd(dx32,fscal,fjx2);
1595             fjy2             = _mm_macc_pd(dy32,fscal,fjy2);
1596             fjz2             = _mm_macc_pd(dz32,fscal,fjz2);
1597
1598             /**************************
1599              * CALCULATE INTERACTIONS *
1600              **************************/
1601
1602             /* REACTION-FIELD ELECTROSTATICS */
1603             felec            = _mm_mul_pd(qq33,_mm_msub_pd(rinv33,rinvsq33,krf2));
1604
1605             fscal            = felec;
1606
1607             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1608
1609             /* Update vectorial force */
1610             fix3             = _mm_macc_pd(dx33,fscal,fix3);
1611             fiy3             = _mm_macc_pd(dy33,fscal,fiy3);
1612             fiz3             = _mm_macc_pd(dz33,fscal,fiz3);
1613
1614             fjx3             = _mm_macc_pd(dx33,fscal,fjx3);
1615             fjy3             = _mm_macc_pd(dy33,fscal,fjy3);
1616             fjz3             = _mm_macc_pd(dz33,fscal,fjz3);
1617
1618             gmx_mm_decrement_4rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2,fjx3,fjy3,fjz3);
1619
1620             /* Inner loop uses 303 flops */
1621         }
1622
1623         /* End of innermost loop */
1624
1625         gmx_mm_update_iforce_4atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,fix3,fiy3,fiz3,
1626                                               f+i_coord_offset,fshift+i_shift_offset);
1627
1628         /* Increment number of inner iterations */
1629         inneriter                  += j_index_end - j_index_start;
1630
1631         /* Outer loop uses 24 flops */
1632     }
1633
1634     /* Increment number of outer iterations */
1635     outeriter        += nri;
1636
1637     /* Update outer/inner flops */
1638
1639     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W4W4_F,outeriter*24 + inneriter*303);
1640 }