src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_double/nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_avx_128_fma_double.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014,2015,2017, 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/gmxlib/nrnb.h"
  46
  47 #include "kernelutil_x86_avx_128_fma_double.h"
  48
  49 /*
  50  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_VF_avx_128_fma_double
  51  * Electrostatics interaction: ReactionField
  52  * VdW interaction:            LennardJones
  53  * Geometry:                   Water3-Water3
  54  * Calculate force/pot:        PotentialAndForce
  55  */
  56 void
  57 nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_VF_avx_128_fma_double
  58                     (t_nblist                    * gmx_restrict       nlist,
  59                      rvec                        * gmx_restrict          xx,
  60                      rvec                        * gmx_restrict          ff,
  61                      struct t_forcerec           * gmx_restrict          fr,
  62                      t_mdatoms                   * gmx_restrict     mdatoms,
  63                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
  64                      t_nrnb                      * gmx_restrict        nrnb)
  65 {
  66     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
  67      * just 0 for non-waters.
  68      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
  69      * jnr indices corresponding to data put in the four positions in the SIMD register.
  70      */
  71     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
  72     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
  73     int              jnrA,jnrB;
  74     int              j_coord_offsetA,j_coord_offsetB;
  75     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
  76     real             rcutoff_scalar;
  77     real             *shiftvec,*fshift,*x,*f;
  78     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
  79     int              vdwioffset0;
  80     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
  81     int              vdwioffset1;
  82     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
  83     int              vdwioffset2;
  84     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
  85     int              vdwjidx0A,vdwjidx0B;
  86     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  87     int              vdwjidx1A,vdwjidx1B;
  88     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  89     int              vdwjidx2A,vdwjidx2B;
  90     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  91     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  92     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  93     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  94     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  95     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  96     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  97     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  98     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
  99     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 100     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 101     real             *charge;
 102     int              nvdwtype;
 103     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 104     int              *vdwtype;
 105     real             *vdwparam;
 106     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 107     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 108     __m128d          dummy_mask,cutoff_mask;
 109     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 110     __m128d          one     = _mm_set1_pd(1.0);
 111     __m128d          two     = _mm_set1_pd(2.0);
 112     x                = xx[0];
 113     f                = ff[0];
 114
 115     nri              = nlist->nri;
 116     iinr             = nlist->iinr;
 117     jindex           = nlist->jindex;
 118     jjnr             = nlist->jjnr;
 119     shiftidx         = nlist->shift;
 120     gid              = nlist->gid;
 121     shiftvec         = fr->shift_vec[0];
 122     fshift           = fr->fshift[0];
 123     facel            = _mm_set1_pd(fr->ic->epsfac);
 124     charge           = mdatoms->chargeA;
 125     krf              = _mm_set1_pd(fr->ic->k_rf);
 126     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 127     crf              = _mm_set1_pd(fr->ic->c_rf);
 128     nvdwtype         = fr->ntype;
 129     vdwparam         = fr->nbfp;
 130     vdwtype          = mdatoms->typeA;
 131
 132     /* Setup water-specific parameters */
 133     inr              = nlist->iinr[0];
 134     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 135     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 136     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 137     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 138
 139     jq0              = _mm_set1_pd(charge[inr+0]);
 140     jq1              = _mm_set1_pd(charge[inr+1]);
 141     jq2              = _mm_set1_pd(charge[inr+2]);
 142     vdwjidx0A        = 2*vdwtype[inr+0];
 143     qq00             = _mm_mul_pd(iq0,jq0);
 144     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 145     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 146     qq01             = _mm_mul_pd(iq0,jq1);
 147     qq02             = _mm_mul_pd(iq0,jq2);
 148     qq10             = _mm_mul_pd(iq1,jq0);
 149     qq11             = _mm_mul_pd(iq1,jq1);
 150     qq12             = _mm_mul_pd(iq1,jq2);
 151     qq20             = _mm_mul_pd(iq2,jq0);
 152     qq21             = _mm_mul_pd(iq2,jq1);
 153     qq22             = _mm_mul_pd(iq2,jq2);
 154
 155     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 156     rcutoff_scalar   = fr->ic->rcoulomb;
 157     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 158     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 159
 160     sh_vdw_invrcut6  = _mm_set1_pd(fr->ic->sh_invrc6);
 161     rvdw             = _mm_set1_pd(fr->ic->rvdw);
 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_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 187                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 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
 199         /* Reset potential sums */
 200         velecsum         = _mm_setzero_pd();
 201         vvdwsum          = _mm_setzero_pd();
 202
 203         /* Start inner kernel loop */
 204         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
 205         {
 206
 207             /* Get j neighbor index, and coordinate index */
 208             jnrA             = jjnr[jidx];
 209             jnrB             = jjnr[jidx+1];
 210             j_coord_offsetA  = DIM*jnrA;
 211             j_coord_offsetB  = DIM*jnrB;
 212
 213             /* load j atom coordinates */
 214             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 215                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 216
 217             /* Calculate displacement vector */
 218             dx00             = _mm_sub_pd(ix0,jx0);
 219             dy00             = _mm_sub_pd(iy0,jy0);
 220             dz00             = _mm_sub_pd(iz0,jz0);
 221             dx01             = _mm_sub_pd(ix0,jx1);
 222             dy01             = _mm_sub_pd(iy0,jy1);
 223             dz01             = _mm_sub_pd(iz0,jz1);
 224             dx02             = _mm_sub_pd(ix0,jx2);
 225             dy02             = _mm_sub_pd(iy0,jy2);
 226             dz02             = _mm_sub_pd(iz0,jz2);
 227             dx10             = _mm_sub_pd(ix1,jx0);
 228             dy10             = _mm_sub_pd(iy1,jy0);
 229             dz10             = _mm_sub_pd(iz1,jz0);
 230             dx11             = _mm_sub_pd(ix1,jx1);
 231             dy11             = _mm_sub_pd(iy1,jy1);
 232             dz11             = _mm_sub_pd(iz1,jz1);
 233             dx12             = _mm_sub_pd(ix1,jx2);
 234             dy12             = _mm_sub_pd(iy1,jy2);
 235             dz12             = _mm_sub_pd(iz1,jz2);
 236             dx20             = _mm_sub_pd(ix2,jx0);
 237             dy20             = _mm_sub_pd(iy2,jy0);
 238             dz20             = _mm_sub_pd(iz2,jz0);
 239             dx21             = _mm_sub_pd(ix2,jx1);
 240             dy21             = _mm_sub_pd(iy2,jy1);
 241             dz21             = _mm_sub_pd(iz2,jz1);
 242             dx22             = _mm_sub_pd(ix2,jx2);
 243             dy22             = _mm_sub_pd(iy2,jy2);
 244             dz22             = _mm_sub_pd(iz2,jz2);
 245
 246             /* Calculate squared distance and things based on it */
 247             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 248             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 249             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 250             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 251             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 252             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 253             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 254             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 255             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 256
 257             rinv00           = avx128fma_invsqrt_d(rsq00);
 258             rinv01           = avx128fma_invsqrt_d(rsq01);
 259             rinv02           = avx128fma_invsqrt_d(rsq02);
 260             rinv10           = avx128fma_invsqrt_d(rsq10);
 261             rinv11           = avx128fma_invsqrt_d(rsq11);
 262             rinv12           = avx128fma_invsqrt_d(rsq12);
 263             rinv20           = avx128fma_invsqrt_d(rsq20);
 264             rinv21           = avx128fma_invsqrt_d(rsq21);
 265             rinv22           = avx128fma_invsqrt_d(rsq22);
 266
 267             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 268             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 269             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 270             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 271             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 272             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 273             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 274             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 275             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 276
 277             fjx0             = _mm_setzero_pd();
 278             fjy0             = _mm_setzero_pd();
 279             fjz0             = _mm_setzero_pd();
 280             fjx1             = _mm_setzero_pd();
 281             fjy1             = _mm_setzero_pd();
 282             fjz1             = _mm_setzero_pd();
 283             fjx2             = _mm_setzero_pd();
 284             fjy2             = _mm_setzero_pd();
 285             fjz2             = _mm_setzero_pd();
 286
 287             /**************************
 288              * CALCULATE INTERACTIONS *
 289              **************************/
 290
 291             if (gmx_mm_any_lt(rsq00,rcutoff2))
 292             {
 293
 294             /* REACTION-FIELD ELECTROSTATICS */
 295             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
 296             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 297
 298             /* LENNARD-JONES DISPERSION/REPULSION */
 299
 300             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 301             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 302             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 303             vvdw             = _mm_msub_pd(_mm_nmacc_pd(c12_00,_mm_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 304                                            _mm_mul_pd(_mm_nmacc_pd( c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 305             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 306
 307             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 308
 309             /* Update potential sum for this i atom from the interaction with this j atom. */
 310             velec            = _mm_and_pd(velec,cutoff_mask);
 311             velecsum         = _mm_add_pd(velecsum,velec);
 312             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 313             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 314
 315             fscal            = _mm_add_pd(felec,fvdw);
 316
 317             fscal            = _mm_and_pd(fscal,cutoff_mask);
 318
 319             /* Update vectorial force */
 320             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 321             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 322             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 323
 324             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 325             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 326             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 327
 328             }
 329
 330             /**************************
 331              * CALCULATE INTERACTIONS *
 332              **************************/
 333
 334             if (gmx_mm_any_lt(rsq01,rcutoff2))
 335             {
 336
 337             /* REACTION-FIELD ELECTROSTATICS */
 338             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_macc_pd(krf,rsq01,rinv01),crf));
 339             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
 340
 341             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
 342
 343             /* Update potential sum for this i atom from the interaction with this j atom. */
 344             velec            = _mm_and_pd(velec,cutoff_mask);
 345             velecsum         = _mm_add_pd(velecsum,velec);
 346
 347             fscal            = felec;
 348
 349             fscal            = _mm_and_pd(fscal,cutoff_mask);
 350
 351             /* Update vectorial force */
 352             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 353             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 354             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 355
 356             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 357             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 358             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 359
 360             }
 361
 362             /**************************
 363              * CALCULATE INTERACTIONS *
 364              **************************/
 365
 366             if (gmx_mm_any_lt(rsq02,rcutoff2))
 367             {
 368
 369             /* REACTION-FIELD ELECTROSTATICS */
 370             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_macc_pd(krf,rsq02,rinv02),crf));
 371             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
 372
 373             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
 374
 375             /* Update potential sum for this i atom from the interaction with this j atom. */
 376             velec            = _mm_and_pd(velec,cutoff_mask);
 377             velecsum         = _mm_add_pd(velecsum,velec);
 378
 379             fscal            = felec;
 380
 381             fscal            = _mm_and_pd(fscal,cutoff_mask);
 382
 383             /* Update vectorial force */
 384             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 385             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 386             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 387
 388             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 389             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 390             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 391
 392             }
 393
 394             /**************************
 395              * CALCULATE INTERACTIONS *
 396              **************************/
 397
 398             if (gmx_mm_any_lt(rsq10,rcutoff2))
 399             {
 400
 401             /* REACTION-FIELD ELECTROSTATICS */
 402             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
 403             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 404
 405             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
 406
 407             /* Update potential sum for this i atom from the interaction with this j atom. */
 408             velec            = _mm_and_pd(velec,cutoff_mask);
 409             velecsum         = _mm_add_pd(velecsum,velec);
 410
 411             fscal            = felec;
 412
 413             fscal            = _mm_and_pd(fscal,cutoff_mask);
 414
 415             /* Update vectorial force */
 416             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 417             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 418             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 419
 420             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 421             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 422             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 423
 424             }
 425
 426             /**************************
 427              * CALCULATE INTERACTIONS *
 428              **************************/
 429
 430             if (gmx_mm_any_lt(rsq11,rcutoff2))
 431             {
 432
 433             /* REACTION-FIELD ELECTROSTATICS */
 434             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 435             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 436
 437             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 438
 439             /* Update potential sum for this i atom from the interaction with this j atom. */
 440             velec            = _mm_and_pd(velec,cutoff_mask);
 441             velecsum         = _mm_add_pd(velecsum,velec);
 442
 443             fscal            = felec;
 444
 445             fscal            = _mm_and_pd(fscal,cutoff_mask);
 446
 447             /* Update vectorial force */
 448             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 449             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 450             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 451
 452             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 453             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 454             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 455
 456             }
 457
 458             /**************************
 459              * CALCULATE INTERACTIONS *
 460              **************************/
 461
 462             if (gmx_mm_any_lt(rsq12,rcutoff2))
 463             {
 464
 465             /* REACTION-FIELD ELECTROSTATICS */
 466             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 467             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 468
 469             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 470
 471             /* Update potential sum for this i atom from the interaction with this j atom. */
 472             velec            = _mm_and_pd(velec,cutoff_mask);
 473             velecsum         = _mm_add_pd(velecsum,velec);
 474
 475             fscal            = felec;
 476
 477             fscal            = _mm_and_pd(fscal,cutoff_mask);
 478
 479             /* Update vectorial force */
 480             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 481             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 482             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 483
 484             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 485             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 486             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 487
 488             }
 489
 490             /**************************
 491              * CALCULATE INTERACTIONS *
 492              **************************/
 493
 494             if (gmx_mm_any_lt(rsq20,rcutoff2))
 495             {
 496
 497             /* REACTION-FIELD ELECTROSTATICS */
 498             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
 499             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 500
 501             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
 502
 503             /* Update potential sum for this i atom from the interaction with this j atom. */
 504             velec            = _mm_and_pd(velec,cutoff_mask);
 505             velecsum         = _mm_add_pd(velecsum,velec);
 506
 507             fscal            = felec;
 508
 509             fscal            = _mm_and_pd(fscal,cutoff_mask);
 510
 511             /* Update vectorial force */
 512             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 513             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 514             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 515
 516             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 517             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 518             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 519
 520             }
 521
 522             /**************************
 523              * CALCULATE INTERACTIONS *
 524              **************************/
 525
 526             if (gmx_mm_any_lt(rsq21,rcutoff2))
 527             {
 528
 529             /* REACTION-FIELD ELECTROSTATICS */
 530             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 531             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 532
 533             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 534
 535             /* Update potential sum for this i atom from the interaction with this j atom. */
 536             velec            = _mm_and_pd(velec,cutoff_mask);
 537             velecsum         = _mm_add_pd(velecsum,velec);
 538
 539             fscal            = felec;
 540
 541             fscal            = _mm_and_pd(fscal,cutoff_mask);
 542
 543             /* Update vectorial force */
 544             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 545             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 546             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 547
 548             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 549             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 550             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 551
 552             }
 553
 554             /**************************
 555              * CALCULATE INTERACTIONS *
 556              **************************/
 557
 558             if (gmx_mm_any_lt(rsq22,rcutoff2))
 559             {
 560
 561             /* REACTION-FIELD ELECTROSTATICS */
 562             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 563             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 564
 565             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 566
 567             /* Update potential sum for this i atom from the interaction with this j atom. */
 568             velec            = _mm_and_pd(velec,cutoff_mask);
 569             velecsum         = _mm_add_pd(velecsum,velec);
 570
 571             fscal            = felec;
 572
 573             fscal            = _mm_and_pd(fscal,cutoff_mask);
 574
 575             /* Update vectorial force */
 576             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 577             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 578             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 579
 580             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 581             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 582             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 583
 584             }
 585
 586             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 587
 588             /* Inner loop uses 369 flops */
 589         }
 590
 591         if(jidx<j_index_end)
 592         {
 593
 594             jnrA             = jjnr[jidx];
 595             j_coord_offsetA  = DIM*jnrA;
 596
 597             /* load j atom coordinates */
 598             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 599                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 600
 601             /* Calculate displacement vector */
 602             dx00             = _mm_sub_pd(ix0,jx0);
 603             dy00             = _mm_sub_pd(iy0,jy0);
 604             dz00             = _mm_sub_pd(iz0,jz0);
 605             dx01             = _mm_sub_pd(ix0,jx1);
 606             dy01             = _mm_sub_pd(iy0,jy1);
 607             dz01             = _mm_sub_pd(iz0,jz1);
 608             dx02             = _mm_sub_pd(ix0,jx2);
 609             dy02             = _mm_sub_pd(iy0,jy2);
 610             dz02             = _mm_sub_pd(iz0,jz2);
 611             dx10             = _mm_sub_pd(ix1,jx0);
 612             dy10             = _mm_sub_pd(iy1,jy0);
 613             dz10             = _mm_sub_pd(iz1,jz0);
 614             dx11             = _mm_sub_pd(ix1,jx1);
 615             dy11             = _mm_sub_pd(iy1,jy1);
 616             dz11             = _mm_sub_pd(iz1,jz1);
 617             dx12             = _mm_sub_pd(ix1,jx2);
 618             dy12             = _mm_sub_pd(iy1,jy2);
 619             dz12             = _mm_sub_pd(iz1,jz2);
 620             dx20             = _mm_sub_pd(ix2,jx0);
 621             dy20             = _mm_sub_pd(iy2,jy0);
 622             dz20             = _mm_sub_pd(iz2,jz0);
 623             dx21             = _mm_sub_pd(ix2,jx1);
 624             dy21             = _mm_sub_pd(iy2,jy1);
 625             dz21             = _mm_sub_pd(iz2,jz1);
 626             dx22             = _mm_sub_pd(ix2,jx2);
 627             dy22             = _mm_sub_pd(iy2,jy2);
 628             dz22             = _mm_sub_pd(iz2,jz2);
 629
 630             /* Calculate squared distance and things based on it */
 631             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 632             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 633             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 634             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 635             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 636             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 637             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 638             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 639             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 640
 641             rinv00           = avx128fma_invsqrt_d(rsq00);
 642             rinv01           = avx128fma_invsqrt_d(rsq01);
 643             rinv02           = avx128fma_invsqrt_d(rsq02);
 644             rinv10           = avx128fma_invsqrt_d(rsq10);
 645             rinv11           = avx128fma_invsqrt_d(rsq11);
 646             rinv12           = avx128fma_invsqrt_d(rsq12);
 647             rinv20           = avx128fma_invsqrt_d(rsq20);
 648             rinv21           = avx128fma_invsqrt_d(rsq21);
 649             rinv22           = avx128fma_invsqrt_d(rsq22);
 650
 651             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 652             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 653             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 654             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 655             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 656             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 657             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 658             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 659             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 660
 661             fjx0             = _mm_setzero_pd();
 662             fjy0             = _mm_setzero_pd();
 663             fjz0             = _mm_setzero_pd();
 664             fjx1             = _mm_setzero_pd();
 665             fjy1             = _mm_setzero_pd();
 666             fjz1             = _mm_setzero_pd();
 667             fjx2             = _mm_setzero_pd();
 668             fjy2             = _mm_setzero_pd();
 669             fjz2             = _mm_setzero_pd();
 670
 671             /**************************
 672              * CALCULATE INTERACTIONS *
 673              **************************/
 674
 675             if (gmx_mm_any_lt(rsq00,rcutoff2))
 676             {
 677
 678             /* REACTION-FIELD ELECTROSTATICS */
 679             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
 680             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 681
 682             /* LENNARD-JONES DISPERSION/REPULSION */
 683
 684             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 685             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 686             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 687             vvdw             = _mm_msub_pd(_mm_nmacc_pd(c12_00,_mm_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 688                                            _mm_mul_pd(_mm_nmacc_pd( c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 689             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 690
 691             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 692
 693             /* Update potential sum for this i atom from the interaction with this j atom. */
 694             velec            = _mm_and_pd(velec,cutoff_mask);
 695             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 696             velecsum         = _mm_add_pd(velecsum,velec);
 697             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 698             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 699             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 700
 701             fscal            = _mm_add_pd(felec,fvdw);
 702
 703             fscal            = _mm_and_pd(fscal,cutoff_mask);
 704
 705             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 706
 707             /* Update vectorial force */
 708             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 709             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 710             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 711
 712             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 713             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 714             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 715
 716             }
 717
 718             /**************************
 719              * CALCULATE INTERACTIONS *
 720              **************************/
 721
 722             if (gmx_mm_any_lt(rsq01,rcutoff2))
 723             {
 724
 725             /* REACTION-FIELD ELECTROSTATICS */
 726             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_macc_pd(krf,rsq01,rinv01),crf));
 727             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
 728
 729             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
 730
 731             /* Update potential sum for this i atom from the interaction with this j atom. */
 732             velec            = _mm_and_pd(velec,cutoff_mask);
 733             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 734             velecsum         = _mm_add_pd(velecsum,velec);
 735
 736             fscal            = felec;
 737
 738             fscal            = _mm_and_pd(fscal,cutoff_mask);
 739
 740             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 741
 742             /* Update vectorial force */
 743             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 744             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 745             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 746
 747             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 748             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 749             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 750
 751             }
 752
 753             /**************************
 754              * CALCULATE INTERACTIONS *
 755              **************************/
 756
 757             if (gmx_mm_any_lt(rsq02,rcutoff2))
 758             {
 759
 760             /* REACTION-FIELD ELECTROSTATICS */
 761             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_macc_pd(krf,rsq02,rinv02),crf));
 762             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
 763
 764             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
 765
 766             /* Update potential sum for this i atom from the interaction with this j atom. */
 767             velec            = _mm_and_pd(velec,cutoff_mask);
 768             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 769             velecsum         = _mm_add_pd(velecsum,velec);
 770
 771             fscal            = felec;
 772
 773             fscal            = _mm_and_pd(fscal,cutoff_mask);
 774
 775             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 776
 777             /* Update vectorial force */
 778             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 779             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 780             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 781
 782             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 783             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 784             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 785
 786             }
 787
 788             /**************************
 789              * CALCULATE INTERACTIONS *
 790              **************************/
 791
 792             if (gmx_mm_any_lt(rsq10,rcutoff2))
 793             {
 794
 795             /* REACTION-FIELD ELECTROSTATICS */
 796             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
 797             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 798
 799             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
 800
 801             /* Update potential sum for this i atom from the interaction with this j atom. */
 802             velec            = _mm_and_pd(velec,cutoff_mask);
 803             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 804             velecsum         = _mm_add_pd(velecsum,velec);
 805
 806             fscal            = felec;
 807
 808             fscal            = _mm_and_pd(fscal,cutoff_mask);
 809
 810             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 811
 812             /* Update vectorial force */
 813             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 814             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 815             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 816
 817             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 818             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 819             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 820
 821             }
 822
 823             /**************************
 824              * CALCULATE INTERACTIONS *
 825              **************************/
 826
 827             if (gmx_mm_any_lt(rsq11,rcutoff2))
 828             {
 829
 830             /* REACTION-FIELD ELECTROSTATICS */
 831             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 832             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 833
 834             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 835
 836             /* Update potential sum for this i atom from the interaction with this j atom. */
 837             velec            = _mm_and_pd(velec,cutoff_mask);
 838             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 839             velecsum         = _mm_add_pd(velecsum,velec);
 840
 841             fscal            = felec;
 842
 843             fscal            = _mm_and_pd(fscal,cutoff_mask);
 844
 845             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 846
 847             /* Update vectorial force */
 848             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 849             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 850             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 851
 852             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 853             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 854             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 855
 856             }
 857
 858             /**************************
 859              * CALCULATE INTERACTIONS *
 860              **************************/
 861
 862             if (gmx_mm_any_lt(rsq12,rcutoff2))
 863             {
 864
 865             /* REACTION-FIELD ELECTROSTATICS */
 866             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 867             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 868
 869             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 870
 871             /* Update potential sum for this i atom from the interaction with this j atom. */
 872             velec            = _mm_and_pd(velec,cutoff_mask);
 873             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 874             velecsum         = _mm_add_pd(velecsum,velec);
 875
 876             fscal            = felec;
 877
 878             fscal            = _mm_and_pd(fscal,cutoff_mask);
 879
 880             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 881
 882             /* Update vectorial force */
 883             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 884             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 885             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 886
 887             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 888             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 889             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 890
 891             }
 892
 893             /**************************
 894              * CALCULATE INTERACTIONS *
 895              **************************/
 896
 897             if (gmx_mm_any_lt(rsq20,rcutoff2))
 898             {
 899
 900             /* REACTION-FIELD ELECTROSTATICS */
 901             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
 902             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 903
 904             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
 905
 906             /* Update potential sum for this i atom from the interaction with this j atom. */
 907             velec            = _mm_and_pd(velec,cutoff_mask);
 908             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 909             velecsum         = _mm_add_pd(velecsum,velec);
 910
 911             fscal            = felec;
 912
 913             fscal            = _mm_and_pd(fscal,cutoff_mask);
 914
 915             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 916
 917             /* Update vectorial force */
 918             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 919             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 920             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 921
 922             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 923             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 924             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 925
 926             }
 927
 928             /**************************
 929              * CALCULATE INTERACTIONS *
 930              **************************/
 931
 932             if (gmx_mm_any_lt(rsq21,rcutoff2))
 933             {
 934
 935             /* REACTION-FIELD ELECTROSTATICS */
 936             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 937             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 938
 939             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 940
 941             /* Update potential sum for this i atom from the interaction with this j atom. */
 942             velec            = _mm_and_pd(velec,cutoff_mask);
 943             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 944             velecsum         = _mm_add_pd(velecsum,velec);
 945
 946             fscal            = felec;
 947
 948             fscal            = _mm_and_pd(fscal,cutoff_mask);
 949
 950             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 951
 952             /* Update vectorial force */
 953             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 954             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 955             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 956
 957             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 958             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 959             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 960
 961             }
 962
 963             /**************************
 964              * CALCULATE INTERACTIONS *
 965              **************************/
 966
 967             if (gmx_mm_any_lt(rsq22,rcutoff2))
 968             {
 969
 970             /* REACTION-FIELD ELECTROSTATICS */
 971             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 972             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 973
 974             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 975
 976             /* Update potential sum for this i atom from the interaction with this j atom. */
 977             velec            = _mm_and_pd(velec,cutoff_mask);
 978             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 979             velecsum         = _mm_add_pd(velecsum,velec);
 980
 981             fscal            = felec;
 982
 983             fscal            = _mm_and_pd(fscal,cutoff_mask);
 984
 985             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 986
 987             /* Update vectorial force */
 988             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 989             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 990             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 991
 992             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 993             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 994             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 995
 996             }
 997
 998             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 999
1000             /* Inner loop uses 369 flops */
1001         }
1002
1003         /* End of innermost loop */
1004
1005         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1006                                               f+i_coord_offset,fshift+i_shift_offset);
1007
1008         ggid                        = gid[iidx];
1009         /* Update potential energies */
1010         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1011         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
1012
1013         /* Increment number of inner iterations */
1014         inneriter                  += j_index_end - j_index_start;
1015
1016         /* Outer loop uses 20 flops */
1017     }
1018
1019     /* Increment number of outer iterations */
1020     outeriter        += nri;
1021
1022     /* Update outer/inner flops */
1023
1024     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*369);
1025 }
1026 /*
1027  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_F_avx_128_fma_double
1028  * Electrostatics interaction: ReactionField
1029  * VdW interaction:            LennardJones
1030  * Geometry:                   Water3-Water3
1031  * Calculate force/pot:        Force
1032  */
1033 void
1034 nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_F_avx_128_fma_double
1035                     (t_nblist                    * gmx_restrict       nlist,
1036                      rvec                        * gmx_restrict          xx,
1037                      rvec                        * gmx_restrict          ff,
1038                      struct t_forcerec           * gmx_restrict          fr,
1039                      t_mdatoms                   * gmx_restrict     mdatoms,
1040                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1041                      t_nrnb                      * gmx_restrict        nrnb)
1042 {
1043     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1044      * just 0 for non-waters.
1045      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1046      * jnr indices corresponding to data put in the four positions in the SIMD register.
1047      */
1048     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1049     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1050     int              jnrA,jnrB;
1051     int              j_coord_offsetA,j_coord_offsetB;
1052     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1053     real             rcutoff_scalar;
1054     real             *shiftvec,*fshift,*x,*f;
1055     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1056     int              vdwioffset0;
1057     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1058     int              vdwioffset1;
1059     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1060     int              vdwioffset2;
1061     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1062     int              vdwjidx0A,vdwjidx0B;
1063     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1064     int              vdwjidx1A,vdwjidx1B;
1065     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1066     int              vdwjidx2A,vdwjidx2B;
1067     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1068     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1069     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1070     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1071     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1072     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1073     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1074     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1075     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1076     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1077     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1078     real             *charge;
1079     int              nvdwtype;
1080     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1081     int              *vdwtype;
1082     real             *vdwparam;
1083     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1084     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1085     __m128d          dummy_mask,cutoff_mask;
1086     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1087     __m128d          one     = _mm_set1_pd(1.0);
1088     __m128d          two     = _mm_set1_pd(2.0);
1089     x                = xx[0];
1090     f                = ff[0];
1091
1092     nri              = nlist->nri;
1093     iinr             = nlist->iinr;
1094     jindex           = nlist->jindex;
1095     jjnr             = nlist->jjnr;
1096     shiftidx         = nlist->shift;
1097     gid              = nlist->gid;
1098     shiftvec         = fr->shift_vec[0];
1099     fshift           = fr->fshift[0];
1100     facel            = _mm_set1_pd(fr->ic->epsfac);
1101     charge           = mdatoms->chargeA;
1102     krf              = _mm_set1_pd(fr->ic->k_rf);
1103     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1104     crf              = _mm_set1_pd(fr->ic->c_rf);
1105     nvdwtype         = fr->ntype;
1106     vdwparam         = fr->nbfp;
1107     vdwtype          = mdatoms->typeA;
1108
1109     /* Setup water-specific parameters */
1110     inr              = nlist->iinr[0];
1111     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
1112     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1113     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1114     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1115
1116     jq0              = _mm_set1_pd(charge[inr+0]);
1117     jq1              = _mm_set1_pd(charge[inr+1]);
1118     jq2              = _mm_set1_pd(charge[inr+2]);
1119     vdwjidx0A        = 2*vdwtype[inr+0];
1120     qq00             = _mm_mul_pd(iq0,jq0);
1121     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1122     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1123     qq01             = _mm_mul_pd(iq0,jq1);
1124     qq02             = _mm_mul_pd(iq0,jq2);
1125     qq10             = _mm_mul_pd(iq1,jq0);
1126     qq11             = _mm_mul_pd(iq1,jq1);
1127     qq12             = _mm_mul_pd(iq1,jq2);
1128     qq20             = _mm_mul_pd(iq2,jq0);
1129     qq21             = _mm_mul_pd(iq2,jq1);
1130     qq22             = _mm_mul_pd(iq2,jq2);
1131
1132     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1133     rcutoff_scalar   = fr->ic->rcoulomb;
1134     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1135     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1136
1137     sh_vdw_invrcut6  = _mm_set1_pd(fr->ic->sh_invrc6);
1138     rvdw             = _mm_set1_pd(fr->ic->rvdw);
1139
1140     /* Avoid stupid compiler warnings */
1141     jnrA = jnrB = 0;
1142     j_coord_offsetA = 0;
1143     j_coord_offsetB = 0;
1144
1145     outeriter        = 0;
1146     inneriter        = 0;
1147
1148     /* Start outer loop over neighborlists */
1149     for(iidx=0; iidx<nri; iidx++)
1150     {
1151         /* Load shift vector for this list */
1152         i_shift_offset   = DIM*shiftidx[iidx];
1153
1154         /* Load limits for loop over neighbors */
1155         j_index_start    = jindex[iidx];
1156         j_index_end      = jindex[iidx+1];
1157
1158         /* Get outer coordinate index */
1159         inr              = iinr[iidx];
1160         i_coord_offset   = DIM*inr;
1161
1162         /* Load i particle coords and add shift vector */
1163         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1164                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1165
1166         fix0             = _mm_setzero_pd();
1167         fiy0             = _mm_setzero_pd();
1168         fiz0             = _mm_setzero_pd();
1169         fix1             = _mm_setzero_pd();
1170         fiy1             = _mm_setzero_pd();
1171         fiz1             = _mm_setzero_pd();
1172         fix2             = _mm_setzero_pd();
1173         fiy2             = _mm_setzero_pd();
1174         fiz2             = _mm_setzero_pd();
1175
1176         /* Start inner kernel loop */
1177         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1178         {
1179
1180             /* Get j neighbor index, and coordinate index */
1181             jnrA             = jjnr[jidx];
1182             jnrB             = jjnr[jidx+1];
1183             j_coord_offsetA  = DIM*jnrA;
1184             j_coord_offsetB  = DIM*jnrB;
1185
1186             /* load j atom coordinates */
1187             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1188                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1189
1190             /* Calculate displacement vector */
1191             dx00             = _mm_sub_pd(ix0,jx0);
1192             dy00             = _mm_sub_pd(iy0,jy0);
1193             dz00             = _mm_sub_pd(iz0,jz0);
1194             dx01             = _mm_sub_pd(ix0,jx1);
1195             dy01             = _mm_sub_pd(iy0,jy1);
1196             dz01             = _mm_sub_pd(iz0,jz1);
1197             dx02             = _mm_sub_pd(ix0,jx2);
1198             dy02             = _mm_sub_pd(iy0,jy2);
1199             dz02             = _mm_sub_pd(iz0,jz2);
1200             dx10             = _mm_sub_pd(ix1,jx0);
1201             dy10             = _mm_sub_pd(iy1,jy0);
1202             dz10             = _mm_sub_pd(iz1,jz0);
1203             dx11             = _mm_sub_pd(ix1,jx1);
1204             dy11             = _mm_sub_pd(iy1,jy1);
1205             dz11             = _mm_sub_pd(iz1,jz1);
1206             dx12             = _mm_sub_pd(ix1,jx2);
1207             dy12             = _mm_sub_pd(iy1,jy2);
1208             dz12             = _mm_sub_pd(iz1,jz2);
1209             dx20             = _mm_sub_pd(ix2,jx0);
1210             dy20             = _mm_sub_pd(iy2,jy0);
1211             dz20             = _mm_sub_pd(iz2,jz0);
1212             dx21             = _mm_sub_pd(ix2,jx1);
1213             dy21             = _mm_sub_pd(iy2,jy1);
1214             dz21             = _mm_sub_pd(iz2,jz1);
1215             dx22             = _mm_sub_pd(ix2,jx2);
1216             dy22             = _mm_sub_pd(iy2,jy2);
1217             dz22             = _mm_sub_pd(iz2,jz2);
1218
1219             /* Calculate squared distance and things based on it */
1220             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1221             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1222             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1223             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1224             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1225             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1226             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1227             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1228             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1229
1230             rinv00           = avx128fma_invsqrt_d(rsq00);
1231             rinv01           = avx128fma_invsqrt_d(rsq01);
1232             rinv02           = avx128fma_invsqrt_d(rsq02);
1233             rinv10           = avx128fma_invsqrt_d(rsq10);
1234             rinv11           = avx128fma_invsqrt_d(rsq11);
1235             rinv12           = avx128fma_invsqrt_d(rsq12);
1236             rinv20           = avx128fma_invsqrt_d(rsq20);
1237             rinv21           = avx128fma_invsqrt_d(rsq21);
1238             rinv22           = avx128fma_invsqrt_d(rsq22);
1239
1240             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1241             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1242             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1243             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1244             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1245             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1246             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1247             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1248             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1249
1250             fjx0             = _mm_setzero_pd();
1251             fjy0             = _mm_setzero_pd();
1252             fjz0             = _mm_setzero_pd();
1253             fjx1             = _mm_setzero_pd();
1254             fjy1             = _mm_setzero_pd();
1255             fjz1             = _mm_setzero_pd();
1256             fjx2             = _mm_setzero_pd();
1257             fjy2             = _mm_setzero_pd();
1258             fjz2             = _mm_setzero_pd();
1259
1260             /**************************
1261              * CALCULATE INTERACTIONS *
1262              **************************/
1263
1264             if (gmx_mm_any_lt(rsq00,rcutoff2))
1265             {
1266
1267             /* REACTION-FIELD ELECTROSTATICS */
1268             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
1269
1270             /* LENNARD-JONES DISPERSION/REPULSION */
1271
1272             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1273             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1274
1275             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1276
1277             fscal            = _mm_add_pd(felec,fvdw);
1278
1279             fscal            = _mm_and_pd(fscal,cutoff_mask);
1280
1281             /* Update vectorial force */
1282             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1283             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1284             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1285
1286             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1287             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1288             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1289
1290             }
1291
1292             /**************************
1293              * CALCULATE INTERACTIONS *
1294              **************************/
1295
1296             if (gmx_mm_any_lt(rsq01,rcutoff2))
1297             {
1298
1299             /* REACTION-FIELD ELECTROSTATICS */
1300             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
1301
1302             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
1303
1304             fscal            = felec;
1305
1306             fscal            = _mm_and_pd(fscal,cutoff_mask);
1307
1308             /* Update vectorial force */
1309             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1310             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1311             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1312
1313             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1314             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1315             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1316
1317             }
1318
1319             /**************************
1320              * CALCULATE INTERACTIONS *
1321              **************************/
1322
1323             if (gmx_mm_any_lt(rsq02,rcutoff2))
1324             {
1325
1326             /* REACTION-FIELD ELECTROSTATICS */
1327             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
1328
1329             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
1330
1331             fscal            = felec;
1332
1333             fscal            = _mm_and_pd(fscal,cutoff_mask);
1334
1335             /* Update vectorial force */
1336             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1337             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1338             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1339
1340             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1341             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1342             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1343
1344             }
1345
1346             /**************************
1347              * CALCULATE INTERACTIONS *
1348              **************************/
1349
1350             if (gmx_mm_any_lt(rsq10,rcutoff2))
1351             {
1352
1353             /* REACTION-FIELD ELECTROSTATICS */
1354             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
1355
1356             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
1357
1358             fscal            = felec;
1359
1360             fscal            = _mm_and_pd(fscal,cutoff_mask);
1361
1362             /* Update vectorial force */
1363             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1364             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1365             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1366
1367             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1368             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1369             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1370
1371             }
1372
1373             /**************************
1374              * CALCULATE INTERACTIONS *
1375              **************************/
1376
1377             if (gmx_mm_any_lt(rsq11,rcutoff2))
1378             {
1379
1380             /* REACTION-FIELD ELECTROSTATICS */
1381             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1382
1383             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1384
1385             fscal            = felec;
1386
1387             fscal            = _mm_and_pd(fscal,cutoff_mask);
1388
1389             /* Update vectorial force */
1390             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1391             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1392             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1393
1394             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1395             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1396             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1397
1398             }
1399
1400             /**************************
1401              * CALCULATE INTERACTIONS *
1402              **************************/
1403
1404             if (gmx_mm_any_lt(rsq12,rcutoff2))
1405             {
1406
1407             /* REACTION-FIELD ELECTROSTATICS */
1408             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1409
1410             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1411
1412             fscal            = felec;
1413
1414             fscal            = _mm_and_pd(fscal,cutoff_mask);
1415
1416             /* Update vectorial force */
1417             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1418             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1419             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1420
1421             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1422             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1423             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1424
1425             }
1426
1427             /**************************
1428              * CALCULATE INTERACTIONS *
1429              **************************/
1430
1431             if (gmx_mm_any_lt(rsq20,rcutoff2))
1432             {
1433
1434             /* REACTION-FIELD ELECTROSTATICS */
1435             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
1436
1437             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
1438
1439             fscal            = felec;
1440
1441             fscal            = _mm_and_pd(fscal,cutoff_mask);
1442
1443             /* Update vectorial force */
1444             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1445             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1446             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1447
1448             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1449             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1450             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1451
1452             }
1453
1454             /**************************
1455              * CALCULATE INTERACTIONS *
1456              **************************/
1457
1458             if (gmx_mm_any_lt(rsq21,rcutoff2))
1459             {
1460
1461             /* REACTION-FIELD ELECTROSTATICS */
1462             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1463
1464             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1465
1466             fscal            = felec;
1467
1468             fscal            = _mm_and_pd(fscal,cutoff_mask);
1469
1470             /* Update vectorial force */
1471             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1472             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1473             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1474
1475             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1476             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1477             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1478
1479             }
1480
1481             /**************************
1482              * CALCULATE INTERACTIONS *
1483              **************************/
1484
1485             if (gmx_mm_any_lt(rsq22,rcutoff2))
1486             {
1487
1488             /* REACTION-FIELD ELECTROSTATICS */
1489             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1490
1491             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1492
1493             fscal            = felec;
1494
1495             fscal            = _mm_and_pd(fscal,cutoff_mask);
1496
1497             /* Update vectorial force */
1498             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1499             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1500             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1501
1502             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1503             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1504             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1505
1506             }
1507
1508             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1509
1510             /* Inner loop uses 304 flops */
1511         }
1512
1513         if(jidx<j_index_end)
1514         {
1515
1516             jnrA             = jjnr[jidx];
1517             j_coord_offsetA  = DIM*jnrA;
1518
1519             /* load j atom coordinates */
1520             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1521                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1522
1523             /* Calculate displacement vector */
1524             dx00             = _mm_sub_pd(ix0,jx0);
1525             dy00             = _mm_sub_pd(iy0,jy0);
1526             dz00             = _mm_sub_pd(iz0,jz0);
1527             dx01             = _mm_sub_pd(ix0,jx1);
1528             dy01             = _mm_sub_pd(iy0,jy1);
1529             dz01             = _mm_sub_pd(iz0,jz1);
1530             dx02             = _mm_sub_pd(ix0,jx2);
1531             dy02             = _mm_sub_pd(iy0,jy2);
1532             dz02             = _mm_sub_pd(iz0,jz2);
1533             dx10             = _mm_sub_pd(ix1,jx0);
1534             dy10             = _mm_sub_pd(iy1,jy0);
1535             dz10             = _mm_sub_pd(iz1,jz0);
1536             dx11             = _mm_sub_pd(ix1,jx1);
1537             dy11             = _mm_sub_pd(iy1,jy1);
1538             dz11             = _mm_sub_pd(iz1,jz1);
1539             dx12             = _mm_sub_pd(ix1,jx2);
1540             dy12             = _mm_sub_pd(iy1,jy2);
1541             dz12             = _mm_sub_pd(iz1,jz2);
1542             dx20             = _mm_sub_pd(ix2,jx0);
1543             dy20             = _mm_sub_pd(iy2,jy0);
1544             dz20             = _mm_sub_pd(iz2,jz0);
1545             dx21             = _mm_sub_pd(ix2,jx1);
1546             dy21             = _mm_sub_pd(iy2,jy1);
1547             dz21             = _mm_sub_pd(iz2,jz1);
1548             dx22             = _mm_sub_pd(ix2,jx2);
1549             dy22             = _mm_sub_pd(iy2,jy2);
1550             dz22             = _mm_sub_pd(iz2,jz2);
1551
1552             /* Calculate squared distance and things based on it */
1553             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1554             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1555             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1556             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1557             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1558             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1559             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1560             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1561             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1562
1563             rinv00           = avx128fma_invsqrt_d(rsq00);
1564             rinv01           = avx128fma_invsqrt_d(rsq01);
1565             rinv02           = avx128fma_invsqrt_d(rsq02);
1566             rinv10           = avx128fma_invsqrt_d(rsq10);
1567             rinv11           = avx128fma_invsqrt_d(rsq11);
1568             rinv12           = avx128fma_invsqrt_d(rsq12);
1569             rinv20           = avx128fma_invsqrt_d(rsq20);
1570             rinv21           = avx128fma_invsqrt_d(rsq21);
1571             rinv22           = avx128fma_invsqrt_d(rsq22);
1572
1573             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1574             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1575             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1576             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1577             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1578             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1579             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1580             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1581             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1582
1583             fjx0             = _mm_setzero_pd();
1584             fjy0             = _mm_setzero_pd();
1585             fjz0             = _mm_setzero_pd();
1586             fjx1             = _mm_setzero_pd();
1587             fjy1             = _mm_setzero_pd();
1588             fjz1             = _mm_setzero_pd();
1589             fjx2             = _mm_setzero_pd();
1590             fjy2             = _mm_setzero_pd();
1591             fjz2             = _mm_setzero_pd();
1592
1593             /**************************
1594              * CALCULATE INTERACTIONS *
1595              **************************/
1596
1597             if (gmx_mm_any_lt(rsq00,rcutoff2))
1598             {
1599
1600             /* REACTION-FIELD ELECTROSTATICS */
1601             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
1602
1603             /* LENNARD-JONES DISPERSION/REPULSION */
1604
1605             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1606             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1607
1608             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1609
1610             fscal            = _mm_add_pd(felec,fvdw);
1611
1612             fscal            = _mm_and_pd(fscal,cutoff_mask);
1613
1614             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1615
1616             /* Update vectorial force */
1617             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1618             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1619             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1620
1621             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1622             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1623             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1624
1625             }
1626
1627             /**************************
1628              * CALCULATE INTERACTIONS *
1629              **************************/
1630
1631             if (gmx_mm_any_lt(rsq01,rcutoff2))
1632             {
1633
1634             /* REACTION-FIELD ELECTROSTATICS */
1635             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
1636
1637             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
1638
1639             fscal            = felec;
1640
1641             fscal            = _mm_and_pd(fscal,cutoff_mask);
1642
1643             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1644
1645             /* Update vectorial force */
1646             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1647             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1648             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1649
1650             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1651             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1652             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1653
1654             }
1655
1656             /**************************
1657              * CALCULATE INTERACTIONS *
1658              **************************/
1659
1660             if (gmx_mm_any_lt(rsq02,rcutoff2))
1661             {
1662
1663             /* REACTION-FIELD ELECTROSTATICS */
1664             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
1665
1666             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
1667
1668             fscal            = felec;
1669
1670             fscal            = _mm_and_pd(fscal,cutoff_mask);
1671
1672             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1673
1674             /* Update vectorial force */
1675             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1676             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1677             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1678
1679             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1680             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1681             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1682
1683             }
1684
1685             /**************************
1686              * CALCULATE INTERACTIONS *
1687              **************************/
1688
1689             if (gmx_mm_any_lt(rsq10,rcutoff2))
1690             {
1691
1692             /* REACTION-FIELD ELECTROSTATICS */
1693             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
1694
1695             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
1696
1697             fscal            = felec;
1698
1699             fscal            = _mm_and_pd(fscal,cutoff_mask);
1700
1701             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1702
1703             /* Update vectorial force */
1704             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1705             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1706             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1707
1708             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1709             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1710             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1711
1712             }
1713
1714             /**************************
1715              * CALCULATE INTERACTIONS *
1716              **************************/
1717
1718             if (gmx_mm_any_lt(rsq11,rcutoff2))
1719             {
1720
1721             /* REACTION-FIELD ELECTROSTATICS */
1722             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1723
1724             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1725
1726             fscal            = felec;
1727
1728             fscal            = _mm_and_pd(fscal,cutoff_mask);
1729
1730             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1731
1732             /* Update vectorial force */
1733             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1734             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1735             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1736
1737             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1738             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1739             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1740
1741             }
1742
1743             /**************************
1744              * CALCULATE INTERACTIONS *
1745              **************************/
1746
1747             if (gmx_mm_any_lt(rsq12,rcutoff2))
1748             {
1749
1750             /* REACTION-FIELD ELECTROSTATICS */
1751             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1752
1753             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1754
1755             fscal            = felec;
1756
1757             fscal            = _mm_and_pd(fscal,cutoff_mask);
1758
1759             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1760
1761             /* Update vectorial force */
1762             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1763             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1764             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1765
1766             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1767             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1768             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1769
1770             }
1771
1772             /**************************
1773              * CALCULATE INTERACTIONS *
1774              **************************/
1775
1776             if (gmx_mm_any_lt(rsq20,rcutoff2))
1777             {
1778
1779             /* REACTION-FIELD ELECTROSTATICS */
1780             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
1781
1782             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
1783
1784             fscal            = felec;
1785
1786             fscal            = _mm_and_pd(fscal,cutoff_mask);
1787
1788             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1789
1790             /* Update vectorial force */
1791             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1792             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1793             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1794
1795             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1796             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1797             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1798
1799             }
1800
1801             /**************************
1802              * CALCULATE INTERACTIONS *
1803              **************************/
1804
1805             if (gmx_mm_any_lt(rsq21,rcutoff2))
1806             {
1807
1808             /* REACTION-FIELD ELECTROSTATICS */
1809             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1810
1811             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1812
1813             fscal            = felec;
1814
1815             fscal            = _mm_and_pd(fscal,cutoff_mask);
1816
1817             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1818
1819             /* Update vectorial force */
1820             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1821             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1822             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1823
1824             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1825             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1826             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1827
1828             }
1829
1830             /**************************
1831              * CALCULATE INTERACTIONS *
1832              **************************/
1833
1834             if (gmx_mm_any_lt(rsq22,rcutoff2))
1835             {
1836
1837             /* REACTION-FIELD ELECTROSTATICS */
1838             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1839
1840             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1841
1842             fscal            = felec;
1843
1844             fscal            = _mm_and_pd(fscal,cutoff_mask);
1845
1846             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1847
1848             /* Update vectorial force */
1849             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1850             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1851             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1852
1853             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1854             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1855             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1856
1857             }
1858
1859             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1860
1861             /* Inner loop uses 304 flops */
1862         }
1863
1864         /* End of innermost loop */
1865
1866         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1867                                               f+i_coord_offset,fshift+i_shift_offset);
1868
1869         /* Increment number of inner iterations */
1870         inneriter                  += j_index_end - j_index_start;
1871
1872         /* Outer loop uses 18 flops */
1873     }
1874
1875     /* Increment number of outer iterations */
1876     outeriter        += nri;
1877
1878     /* Update outer/inner flops */
1879
1880     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*304);
1881 }