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, 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 "gromacs/legacyheaders/types/simple.h"
  44 #include "gromacs/math/vec.h"
  45 #include "gromacs/legacyheaders/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_ElecRFCut_VdwLJSh_GeomW3W3_VF_avx_128_fma_double
  52  * Electrostatics interaction: ReactionField
  53  * VdW interaction:            LennardJones
  54  * Geometry:                   Water3-Water3
  55  * Calculate force/pot:        PotentialAndForce
  56  */
  57 void
  58 nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_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              vdwjidx0A,vdwjidx0B;
  87     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
  88     int              vdwjidx1A,vdwjidx1B;
  89     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
  90     int              vdwjidx2A,vdwjidx2B;
  91     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
  92     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
  93     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
  94     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
  95     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
  96     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
  97     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
  98     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
  99     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
 100     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
 101     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
 102     real             *charge;
 103     int              nvdwtype;
 104     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
 105     int              *vdwtype;
 106     real             *vdwparam;
 107     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
 108     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
 109     __m128d          dummy_mask,cutoff_mask;
 110     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
 111     __m128d          one     = _mm_set1_pd(1.0);
 112     __m128d          two     = _mm_set1_pd(2.0);
 113     x                = xx[0];
 114     f                = ff[0];
 115
 116     nri              = nlist->nri;
 117     iinr             = nlist->iinr;
 118     jindex           = nlist->jindex;
 119     jjnr             = nlist->jjnr;
 120     shiftidx         = nlist->shift;
 121     gid              = nlist->gid;
 122     shiftvec         = fr->shift_vec[0];
 123     fshift           = fr->fshift[0];
 124     facel            = _mm_set1_pd(fr->epsfac);
 125     charge           = mdatoms->chargeA;
 126     krf              = _mm_set1_pd(fr->ic->k_rf);
 127     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
 128     crf              = _mm_set1_pd(fr->ic->c_rf);
 129     nvdwtype         = fr->ntype;
 130     vdwparam         = fr->nbfp;
 131     vdwtype          = mdatoms->typeA;
 132
 133     /* Setup water-specific parameters */
 134     inr              = nlist->iinr[0];
 135     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
 136     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
 137     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
 138     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
 139
 140     jq0              = _mm_set1_pd(charge[inr+0]);
 141     jq1              = _mm_set1_pd(charge[inr+1]);
 142     jq2              = _mm_set1_pd(charge[inr+2]);
 143     vdwjidx0A        = 2*vdwtype[inr+0];
 144     qq00             = _mm_mul_pd(iq0,jq0);
 145     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
 146     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
 147     qq01             = _mm_mul_pd(iq0,jq1);
 148     qq02             = _mm_mul_pd(iq0,jq2);
 149     qq10             = _mm_mul_pd(iq1,jq0);
 150     qq11             = _mm_mul_pd(iq1,jq1);
 151     qq12             = _mm_mul_pd(iq1,jq2);
 152     qq20             = _mm_mul_pd(iq2,jq0);
 153     qq21             = _mm_mul_pd(iq2,jq1);
 154     qq22             = _mm_mul_pd(iq2,jq2);
 155
 156     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
 157     rcutoff_scalar   = fr->rcoulomb;
 158     rcutoff          = _mm_set1_pd(rcutoff_scalar);
 159     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
 160
 161     sh_vdw_invrcut6  = _mm_set1_pd(fr->ic->sh_invrc6);
 162     rvdw             = _mm_set1_pd(fr->rvdw);
 163
 164     /* Avoid stupid compiler warnings */
 165     jnrA = jnrB = 0;
 166     j_coord_offsetA = 0;
 167     j_coord_offsetB = 0;
 168
 169     outeriter        = 0;
 170     inneriter        = 0;
 171
 172     /* Start outer loop over neighborlists */
 173     for(iidx=0; iidx<nri; iidx++)
 174     {
 175         /* Load shift vector for this list */
 176         i_shift_offset   = DIM*shiftidx[iidx];
 177
 178         /* Load limits for loop over neighbors */
 179         j_index_start    = jindex[iidx];
 180         j_index_end      = jindex[iidx+1];
 181
 182         /* Get outer coordinate index */
 183         inr              = iinr[iidx];
 184         i_coord_offset   = DIM*inr;
 185
 186         /* Load i particle coords and add shift vector */
 187         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
 188                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
 189
 190         fix0             = _mm_setzero_pd();
 191         fiy0             = _mm_setzero_pd();
 192         fiz0             = _mm_setzero_pd();
 193         fix1             = _mm_setzero_pd();
 194         fiy1             = _mm_setzero_pd();
 195         fiz1             = _mm_setzero_pd();
 196         fix2             = _mm_setzero_pd();
 197         fiy2             = _mm_setzero_pd();
 198         fiz2             = _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_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
 216                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 217
 218             /* Calculate displacement vector */
 219             dx00             = _mm_sub_pd(ix0,jx0);
 220             dy00             = _mm_sub_pd(iy0,jy0);
 221             dz00             = _mm_sub_pd(iz0,jz0);
 222             dx01             = _mm_sub_pd(ix0,jx1);
 223             dy01             = _mm_sub_pd(iy0,jy1);
 224             dz01             = _mm_sub_pd(iz0,jz1);
 225             dx02             = _mm_sub_pd(ix0,jx2);
 226             dy02             = _mm_sub_pd(iy0,jy2);
 227             dz02             = _mm_sub_pd(iz0,jz2);
 228             dx10             = _mm_sub_pd(ix1,jx0);
 229             dy10             = _mm_sub_pd(iy1,jy0);
 230             dz10             = _mm_sub_pd(iz1,jz0);
 231             dx11             = _mm_sub_pd(ix1,jx1);
 232             dy11             = _mm_sub_pd(iy1,jy1);
 233             dz11             = _mm_sub_pd(iz1,jz1);
 234             dx12             = _mm_sub_pd(ix1,jx2);
 235             dy12             = _mm_sub_pd(iy1,jy2);
 236             dz12             = _mm_sub_pd(iz1,jz2);
 237             dx20             = _mm_sub_pd(ix2,jx0);
 238             dy20             = _mm_sub_pd(iy2,jy0);
 239             dz20             = _mm_sub_pd(iz2,jz0);
 240             dx21             = _mm_sub_pd(ix2,jx1);
 241             dy21             = _mm_sub_pd(iy2,jy1);
 242             dz21             = _mm_sub_pd(iz2,jz1);
 243             dx22             = _mm_sub_pd(ix2,jx2);
 244             dy22             = _mm_sub_pd(iy2,jy2);
 245             dz22             = _mm_sub_pd(iz2,jz2);
 246
 247             /* Calculate squared distance and things based on it */
 248             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 249             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 250             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 251             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 252             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 253             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 254             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 255             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 256             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 257
 258             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 259             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 260             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 261             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 262             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 263             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 264             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 265             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 266             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 267
 268             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 269             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 270             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 271             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 272             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 273             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 274             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 275             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 276             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 277
 278             fjx0             = _mm_setzero_pd();
 279             fjy0             = _mm_setzero_pd();
 280             fjz0             = _mm_setzero_pd();
 281             fjx1             = _mm_setzero_pd();
 282             fjy1             = _mm_setzero_pd();
 283             fjz1             = _mm_setzero_pd();
 284             fjx2             = _mm_setzero_pd();
 285             fjy2             = _mm_setzero_pd();
 286             fjz2             = _mm_setzero_pd();
 287
 288             /**************************
 289              * CALCULATE INTERACTIONS *
 290              **************************/
 291
 292             if (gmx_mm_any_lt(rsq00,rcutoff2))
 293             {
 294
 295             /* REACTION-FIELD ELECTROSTATICS */
 296             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
 297             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 298
 299             /* LENNARD-JONES DISPERSION/REPULSION */
 300
 301             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 302             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 303             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 304             vvdw             = _mm_msub_pd(_mm_nmacc_pd(c12_00,_mm_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 305                                            _mm_mul_pd(_mm_nmacc_pd( c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 306             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 307
 308             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 309
 310             /* Update potential sum for this i atom from the interaction with this j atom. */
 311             velec            = _mm_and_pd(velec,cutoff_mask);
 312             velecsum         = _mm_add_pd(velecsum,velec);
 313             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 314             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 315
 316             fscal            = _mm_add_pd(felec,fvdw);
 317
 318             fscal            = _mm_and_pd(fscal,cutoff_mask);
 319
 320             /* Update vectorial force */
 321             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 322             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 323             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 324
 325             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 326             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 327             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 328
 329             }
 330
 331             /**************************
 332              * CALCULATE INTERACTIONS *
 333              **************************/
 334
 335             if (gmx_mm_any_lt(rsq01,rcutoff2))
 336             {
 337
 338             /* REACTION-FIELD ELECTROSTATICS */
 339             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_macc_pd(krf,rsq01,rinv01),crf));
 340             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
 341
 342             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
 343
 344             /* Update potential sum for this i atom from the interaction with this j atom. */
 345             velec            = _mm_and_pd(velec,cutoff_mask);
 346             velecsum         = _mm_add_pd(velecsum,velec);
 347
 348             fscal            = felec;
 349
 350             fscal            = _mm_and_pd(fscal,cutoff_mask);
 351
 352             /* Update vectorial force */
 353             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 354             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 355             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 356
 357             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 358             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 359             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 360
 361             }
 362
 363             /**************************
 364              * CALCULATE INTERACTIONS *
 365              **************************/
 366
 367             if (gmx_mm_any_lt(rsq02,rcutoff2))
 368             {
 369
 370             /* REACTION-FIELD ELECTROSTATICS */
 371             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_macc_pd(krf,rsq02,rinv02),crf));
 372             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
 373
 374             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
 375
 376             /* Update potential sum for this i atom from the interaction with this j atom. */
 377             velec            = _mm_and_pd(velec,cutoff_mask);
 378             velecsum         = _mm_add_pd(velecsum,velec);
 379
 380             fscal            = felec;
 381
 382             fscal            = _mm_and_pd(fscal,cutoff_mask);
 383
 384             /* Update vectorial force */
 385             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 386             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 387             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 388
 389             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 390             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 391             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 392
 393             }
 394
 395             /**************************
 396              * CALCULATE INTERACTIONS *
 397              **************************/
 398
 399             if (gmx_mm_any_lt(rsq10,rcutoff2))
 400             {
 401
 402             /* REACTION-FIELD ELECTROSTATICS */
 403             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
 404             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 405
 406             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
 407
 408             /* Update potential sum for this i atom from the interaction with this j atom. */
 409             velec            = _mm_and_pd(velec,cutoff_mask);
 410             velecsum         = _mm_add_pd(velecsum,velec);
 411
 412             fscal            = felec;
 413
 414             fscal            = _mm_and_pd(fscal,cutoff_mask);
 415
 416             /* Update vectorial force */
 417             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 418             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 419             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 420
 421             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 422             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 423             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 424
 425             }
 426
 427             /**************************
 428              * CALCULATE INTERACTIONS *
 429              **************************/
 430
 431             if (gmx_mm_any_lt(rsq11,rcutoff2))
 432             {
 433
 434             /* REACTION-FIELD ELECTROSTATICS */
 435             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 436             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 437
 438             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 439
 440             /* Update potential sum for this i atom from the interaction with this j atom. */
 441             velec            = _mm_and_pd(velec,cutoff_mask);
 442             velecsum         = _mm_add_pd(velecsum,velec);
 443
 444             fscal            = felec;
 445
 446             fscal            = _mm_and_pd(fscal,cutoff_mask);
 447
 448             /* Update vectorial force */
 449             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 450             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 451             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 452
 453             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 454             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 455             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 456
 457             }
 458
 459             /**************************
 460              * CALCULATE INTERACTIONS *
 461              **************************/
 462
 463             if (gmx_mm_any_lt(rsq12,rcutoff2))
 464             {
 465
 466             /* REACTION-FIELD ELECTROSTATICS */
 467             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 468             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 469
 470             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 471
 472             /* Update potential sum for this i atom from the interaction with this j atom. */
 473             velec            = _mm_and_pd(velec,cutoff_mask);
 474             velecsum         = _mm_add_pd(velecsum,velec);
 475
 476             fscal            = felec;
 477
 478             fscal            = _mm_and_pd(fscal,cutoff_mask);
 479
 480             /* Update vectorial force */
 481             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 482             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 483             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 484
 485             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 486             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 487             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 488
 489             }
 490
 491             /**************************
 492              * CALCULATE INTERACTIONS *
 493              **************************/
 494
 495             if (gmx_mm_any_lt(rsq20,rcutoff2))
 496             {
 497
 498             /* REACTION-FIELD ELECTROSTATICS */
 499             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
 500             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 501
 502             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
 503
 504             /* Update potential sum for this i atom from the interaction with this j atom. */
 505             velec            = _mm_and_pd(velec,cutoff_mask);
 506             velecsum         = _mm_add_pd(velecsum,velec);
 507
 508             fscal            = felec;
 509
 510             fscal            = _mm_and_pd(fscal,cutoff_mask);
 511
 512             /* Update vectorial force */
 513             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 514             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 515             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 516
 517             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 518             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 519             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 520
 521             }
 522
 523             /**************************
 524              * CALCULATE INTERACTIONS *
 525              **************************/
 526
 527             if (gmx_mm_any_lt(rsq21,rcutoff2))
 528             {
 529
 530             /* REACTION-FIELD ELECTROSTATICS */
 531             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 532             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 533
 534             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 535
 536             /* Update potential sum for this i atom from the interaction with this j atom. */
 537             velec            = _mm_and_pd(velec,cutoff_mask);
 538             velecsum         = _mm_add_pd(velecsum,velec);
 539
 540             fscal            = felec;
 541
 542             fscal            = _mm_and_pd(fscal,cutoff_mask);
 543
 544             /* Update vectorial force */
 545             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 546             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 547             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 548
 549             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 550             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 551             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 552
 553             }
 554
 555             /**************************
 556              * CALCULATE INTERACTIONS *
 557              **************************/
 558
 559             if (gmx_mm_any_lt(rsq22,rcutoff2))
 560             {
 561
 562             /* REACTION-FIELD ELECTROSTATICS */
 563             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 564             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 565
 566             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 567
 568             /* Update potential sum for this i atom from the interaction with this j atom. */
 569             velec            = _mm_and_pd(velec,cutoff_mask);
 570             velecsum         = _mm_add_pd(velecsum,velec);
 571
 572             fscal            = felec;
 573
 574             fscal            = _mm_and_pd(fscal,cutoff_mask);
 575
 576             /* Update vectorial force */
 577             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 578             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 579             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 580
 581             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 582             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 583             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 584
 585             }
 586
 587             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
 588
 589             /* Inner loop uses 369 flops */
 590         }
 591
 592         if(jidx<j_index_end)
 593         {
 594
 595             jnrA             = jjnr[jidx];
 596             j_coord_offsetA  = DIM*jnrA;
 597
 598             /* load j atom coordinates */
 599             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
 600                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
 601
 602             /* Calculate displacement vector */
 603             dx00             = _mm_sub_pd(ix0,jx0);
 604             dy00             = _mm_sub_pd(iy0,jy0);
 605             dz00             = _mm_sub_pd(iz0,jz0);
 606             dx01             = _mm_sub_pd(ix0,jx1);
 607             dy01             = _mm_sub_pd(iy0,jy1);
 608             dz01             = _mm_sub_pd(iz0,jz1);
 609             dx02             = _mm_sub_pd(ix0,jx2);
 610             dy02             = _mm_sub_pd(iy0,jy2);
 611             dz02             = _mm_sub_pd(iz0,jz2);
 612             dx10             = _mm_sub_pd(ix1,jx0);
 613             dy10             = _mm_sub_pd(iy1,jy0);
 614             dz10             = _mm_sub_pd(iz1,jz0);
 615             dx11             = _mm_sub_pd(ix1,jx1);
 616             dy11             = _mm_sub_pd(iy1,jy1);
 617             dz11             = _mm_sub_pd(iz1,jz1);
 618             dx12             = _mm_sub_pd(ix1,jx2);
 619             dy12             = _mm_sub_pd(iy1,jy2);
 620             dz12             = _mm_sub_pd(iz1,jz2);
 621             dx20             = _mm_sub_pd(ix2,jx0);
 622             dy20             = _mm_sub_pd(iy2,jy0);
 623             dz20             = _mm_sub_pd(iz2,jz0);
 624             dx21             = _mm_sub_pd(ix2,jx1);
 625             dy21             = _mm_sub_pd(iy2,jy1);
 626             dz21             = _mm_sub_pd(iz2,jz1);
 627             dx22             = _mm_sub_pd(ix2,jx2);
 628             dy22             = _mm_sub_pd(iy2,jy2);
 629             dz22             = _mm_sub_pd(iz2,jz2);
 630
 631             /* Calculate squared distance and things based on it */
 632             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
 633             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
 634             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
 635             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
 636             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
 637             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
 638             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
 639             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
 640             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
 641
 642             rinv00           = gmx_mm_invsqrt_pd(rsq00);
 643             rinv01           = gmx_mm_invsqrt_pd(rsq01);
 644             rinv02           = gmx_mm_invsqrt_pd(rsq02);
 645             rinv10           = gmx_mm_invsqrt_pd(rsq10);
 646             rinv11           = gmx_mm_invsqrt_pd(rsq11);
 647             rinv12           = gmx_mm_invsqrt_pd(rsq12);
 648             rinv20           = gmx_mm_invsqrt_pd(rsq20);
 649             rinv21           = gmx_mm_invsqrt_pd(rsq21);
 650             rinv22           = gmx_mm_invsqrt_pd(rsq22);
 651
 652             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
 653             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
 654             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
 655             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
 656             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
 657             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
 658             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
 659             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
 660             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
 661
 662             fjx0             = _mm_setzero_pd();
 663             fjy0             = _mm_setzero_pd();
 664             fjz0             = _mm_setzero_pd();
 665             fjx1             = _mm_setzero_pd();
 666             fjy1             = _mm_setzero_pd();
 667             fjz1             = _mm_setzero_pd();
 668             fjx2             = _mm_setzero_pd();
 669             fjy2             = _mm_setzero_pd();
 670             fjz2             = _mm_setzero_pd();
 671
 672             /**************************
 673              * CALCULATE INTERACTIONS *
 674              **************************/
 675
 676             if (gmx_mm_any_lt(rsq00,rcutoff2))
 677             {
 678
 679             /* REACTION-FIELD ELECTROSTATICS */
 680             velec            = _mm_mul_pd(qq00,_mm_sub_pd(_mm_macc_pd(krf,rsq00,rinv00),crf));
 681             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
 682
 683             /* LENNARD-JONES DISPERSION/REPULSION */
 684
 685             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
 686             vvdw6            = _mm_mul_pd(c6_00,rinvsix);
 687             vvdw12           = _mm_mul_pd(c12_00,_mm_mul_pd(rinvsix,rinvsix));
 688             vvdw             = _mm_msub_pd(_mm_nmacc_pd(c12_00,_mm_mul_pd(sh_vdw_invrcut6,sh_vdw_invrcut6),vvdw12),one_twelfth,
 689                                            _mm_mul_pd(_mm_nmacc_pd( c6_00,sh_vdw_invrcut6,vvdw6),one_sixth));
 690             fvdw             = _mm_mul_pd(_mm_sub_pd(vvdw12,vvdw6),rinvsq00);
 691
 692             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
 693
 694             /* Update potential sum for this i atom from the interaction with this j atom. */
 695             velec            = _mm_and_pd(velec,cutoff_mask);
 696             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 697             velecsum         = _mm_add_pd(velecsum,velec);
 698             vvdw             = _mm_and_pd(vvdw,cutoff_mask);
 699             vvdw             = _mm_unpacklo_pd(vvdw,_mm_setzero_pd());
 700             vvdwsum          = _mm_add_pd(vvdwsum,vvdw);
 701
 702             fscal            = _mm_add_pd(felec,fvdw);
 703
 704             fscal            = _mm_and_pd(fscal,cutoff_mask);
 705
 706             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 707
 708             /* Update vectorial force */
 709             fix0             = _mm_macc_pd(dx00,fscal,fix0);
 710             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
 711             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
 712
 713             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
 714             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
 715             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
 716
 717             }
 718
 719             /**************************
 720              * CALCULATE INTERACTIONS *
 721              **************************/
 722
 723             if (gmx_mm_any_lt(rsq01,rcutoff2))
 724             {
 725
 726             /* REACTION-FIELD ELECTROSTATICS */
 727             velec            = _mm_mul_pd(qq01,_mm_sub_pd(_mm_macc_pd(krf,rsq01,rinv01),crf));
 728             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
 729
 730             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
 731
 732             /* Update potential sum for this i atom from the interaction with this j atom. */
 733             velec            = _mm_and_pd(velec,cutoff_mask);
 734             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 735             velecsum         = _mm_add_pd(velecsum,velec);
 736
 737             fscal            = felec;
 738
 739             fscal            = _mm_and_pd(fscal,cutoff_mask);
 740
 741             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 742
 743             /* Update vectorial force */
 744             fix0             = _mm_macc_pd(dx01,fscal,fix0);
 745             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
 746             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
 747
 748             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
 749             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
 750             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
 751
 752             }
 753
 754             /**************************
 755              * CALCULATE INTERACTIONS *
 756              **************************/
 757
 758             if (gmx_mm_any_lt(rsq02,rcutoff2))
 759             {
 760
 761             /* REACTION-FIELD ELECTROSTATICS */
 762             velec            = _mm_mul_pd(qq02,_mm_sub_pd(_mm_macc_pd(krf,rsq02,rinv02),crf));
 763             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
 764
 765             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
 766
 767             /* Update potential sum for this i atom from the interaction with this j atom. */
 768             velec            = _mm_and_pd(velec,cutoff_mask);
 769             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 770             velecsum         = _mm_add_pd(velecsum,velec);
 771
 772             fscal            = felec;
 773
 774             fscal            = _mm_and_pd(fscal,cutoff_mask);
 775
 776             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 777
 778             /* Update vectorial force */
 779             fix0             = _mm_macc_pd(dx02,fscal,fix0);
 780             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
 781             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
 782
 783             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
 784             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
 785             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
 786
 787             }
 788
 789             /**************************
 790              * CALCULATE INTERACTIONS *
 791              **************************/
 792
 793             if (gmx_mm_any_lt(rsq10,rcutoff2))
 794             {
 795
 796             /* REACTION-FIELD ELECTROSTATICS */
 797             velec            = _mm_mul_pd(qq10,_mm_sub_pd(_mm_macc_pd(krf,rsq10,rinv10),crf));
 798             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
 799
 800             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
 801
 802             /* Update potential sum for this i atom from the interaction with this j atom. */
 803             velec            = _mm_and_pd(velec,cutoff_mask);
 804             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 805             velecsum         = _mm_add_pd(velecsum,velec);
 806
 807             fscal            = felec;
 808
 809             fscal            = _mm_and_pd(fscal,cutoff_mask);
 810
 811             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 812
 813             /* Update vectorial force */
 814             fix1             = _mm_macc_pd(dx10,fscal,fix1);
 815             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
 816             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
 817
 818             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
 819             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
 820             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
 821
 822             }
 823
 824             /**************************
 825              * CALCULATE INTERACTIONS *
 826              **************************/
 827
 828             if (gmx_mm_any_lt(rsq11,rcutoff2))
 829             {
 830
 831             /* REACTION-FIELD ELECTROSTATICS */
 832             velec            = _mm_mul_pd(qq11,_mm_sub_pd(_mm_macc_pd(krf,rsq11,rinv11),crf));
 833             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
 834
 835             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
 836
 837             /* Update potential sum for this i atom from the interaction with this j atom. */
 838             velec            = _mm_and_pd(velec,cutoff_mask);
 839             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 840             velecsum         = _mm_add_pd(velecsum,velec);
 841
 842             fscal            = felec;
 843
 844             fscal            = _mm_and_pd(fscal,cutoff_mask);
 845
 846             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 847
 848             /* Update vectorial force */
 849             fix1             = _mm_macc_pd(dx11,fscal,fix1);
 850             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
 851             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
 852
 853             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
 854             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
 855             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
 856
 857             }
 858
 859             /**************************
 860              * CALCULATE INTERACTIONS *
 861              **************************/
 862
 863             if (gmx_mm_any_lt(rsq12,rcutoff2))
 864             {
 865
 866             /* REACTION-FIELD ELECTROSTATICS */
 867             velec            = _mm_mul_pd(qq12,_mm_sub_pd(_mm_macc_pd(krf,rsq12,rinv12),crf));
 868             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
 869
 870             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
 871
 872             /* Update potential sum for this i atom from the interaction with this j atom. */
 873             velec            = _mm_and_pd(velec,cutoff_mask);
 874             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 875             velecsum         = _mm_add_pd(velecsum,velec);
 876
 877             fscal            = felec;
 878
 879             fscal            = _mm_and_pd(fscal,cutoff_mask);
 880
 881             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 882
 883             /* Update vectorial force */
 884             fix1             = _mm_macc_pd(dx12,fscal,fix1);
 885             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
 886             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
 887
 888             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
 889             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
 890             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
 891
 892             }
 893
 894             /**************************
 895              * CALCULATE INTERACTIONS *
 896              **************************/
 897
 898             if (gmx_mm_any_lt(rsq20,rcutoff2))
 899             {
 900
 901             /* REACTION-FIELD ELECTROSTATICS */
 902             velec            = _mm_mul_pd(qq20,_mm_sub_pd(_mm_macc_pd(krf,rsq20,rinv20),crf));
 903             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
 904
 905             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
 906
 907             /* Update potential sum for this i atom from the interaction with this j atom. */
 908             velec            = _mm_and_pd(velec,cutoff_mask);
 909             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 910             velecsum         = _mm_add_pd(velecsum,velec);
 911
 912             fscal            = felec;
 913
 914             fscal            = _mm_and_pd(fscal,cutoff_mask);
 915
 916             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 917
 918             /* Update vectorial force */
 919             fix2             = _mm_macc_pd(dx20,fscal,fix2);
 920             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
 921             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
 922
 923             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
 924             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
 925             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
 926
 927             }
 928
 929             /**************************
 930              * CALCULATE INTERACTIONS *
 931              **************************/
 932
 933             if (gmx_mm_any_lt(rsq21,rcutoff2))
 934             {
 935
 936             /* REACTION-FIELD ELECTROSTATICS */
 937             velec            = _mm_mul_pd(qq21,_mm_sub_pd(_mm_macc_pd(krf,rsq21,rinv21),crf));
 938             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
 939
 940             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
 941
 942             /* Update potential sum for this i atom from the interaction with this j atom. */
 943             velec            = _mm_and_pd(velec,cutoff_mask);
 944             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 945             velecsum         = _mm_add_pd(velecsum,velec);
 946
 947             fscal            = felec;
 948
 949             fscal            = _mm_and_pd(fscal,cutoff_mask);
 950
 951             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 952
 953             /* Update vectorial force */
 954             fix2             = _mm_macc_pd(dx21,fscal,fix2);
 955             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
 956             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
 957
 958             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
 959             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
 960             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
 961
 962             }
 963
 964             /**************************
 965              * CALCULATE INTERACTIONS *
 966              **************************/
 967
 968             if (gmx_mm_any_lt(rsq22,rcutoff2))
 969             {
 970
 971             /* REACTION-FIELD ELECTROSTATICS */
 972             velec            = _mm_mul_pd(qq22,_mm_sub_pd(_mm_macc_pd(krf,rsq22,rinv22),crf));
 973             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
 974
 975             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
 976
 977             /* Update potential sum for this i atom from the interaction with this j atom. */
 978             velec            = _mm_and_pd(velec,cutoff_mask);
 979             velec            = _mm_unpacklo_pd(velec,_mm_setzero_pd());
 980             velecsum         = _mm_add_pd(velecsum,velec);
 981
 982             fscal            = felec;
 983
 984             fscal            = _mm_and_pd(fscal,cutoff_mask);
 985
 986             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
 987
 988             /* Update vectorial force */
 989             fix2             = _mm_macc_pd(dx22,fscal,fix2);
 990             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
 991             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
 992
 993             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
 994             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
 995             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
 996
 997             }
 998
 999             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1000
1001             /* Inner loop uses 369 flops */
1002         }
1003
1004         /* End of innermost loop */
1005
1006         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1007                                               f+i_coord_offset,fshift+i_shift_offset);
1008
1009         ggid                        = gid[iidx];
1010         /* Update potential energies */
1011         gmx_mm_update_1pot_pd(velecsum,kernel_data->energygrp_elec+ggid);
1012         gmx_mm_update_1pot_pd(vvdwsum,kernel_data->energygrp_vdw+ggid);
1013
1014         /* Increment number of inner iterations */
1015         inneriter                  += j_index_end - j_index_start;
1016
1017         /* Outer loop uses 20 flops */
1018     }
1019
1020     /* Increment number of outer iterations */
1021     outeriter        += nri;
1022
1023     /* Update outer/inner flops */
1024
1025     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_VF,outeriter*20 + inneriter*369);
1026 }
1027 /*
1028  * Gromacs nonbonded kernel:   nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_F_avx_128_fma_double
1029  * Electrostatics interaction: ReactionField
1030  * VdW interaction:            LennardJones
1031  * Geometry:                   Water3-Water3
1032  * Calculate force/pot:        Force
1033  */
1034 void
1035 nb_kernel_ElecRFCut_VdwLJSh_GeomW3W3_F_avx_128_fma_double
1036                     (t_nblist                    * gmx_restrict       nlist,
1037                      rvec                        * gmx_restrict          xx,
1038                      rvec                        * gmx_restrict          ff,
1039                      t_forcerec                  * gmx_restrict          fr,
1040                      t_mdatoms                   * gmx_restrict     mdatoms,
1041                      nb_kernel_data_t gmx_unused * gmx_restrict kernel_data,
1042                      t_nrnb                      * gmx_restrict        nrnb)
1043 {
1044     /* Suffixes 0,1,2,3 refer to particle indices for waters in the inner or outer loop, or
1045      * just 0 for non-waters.
1046      * Suffixes A,B refer to j loop unrolling done with SSE double precision, e.g. for the two different
1047      * jnr indices corresponding to data put in the four positions in the SIMD register.
1048      */
1049     int              i_shift_offset,i_coord_offset,outeriter,inneriter;
1050     int              j_index_start,j_index_end,jidx,nri,inr,ggid,iidx;
1051     int              jnrA,jnrB;
1052     int              j_coord_offsetA,j_coord_offsetB;
1053     int              *iinr,*jindex,*jjnr,*shiftidx,*gid;
1054     real             rcutoff_scalar;
1055     real             *shiftvec,*fshift,*x,*f;
1056     __m128d          tx,ty,tz,fscal,rcutoff,rcutoff2,jidxall;
1057     int              vdwioffset0;
1058     __m128d          ix0,iy0,iz0,fix0,fiy0,fiz0,iq0,isai0;
1059     int              vdwioffset1;
1060     __m128d          ix1,iy1,iz1,fix1,fiy1,fiz1,iq1,isai1;
1061     int              vdwioffset2;
1062     __m128d          ix2,iy2,iz2,fix2,fiy2,fiz2,iq2,isai2;
1063     int              vdwjidx0A,vdwjidx0B;
1064     __m128d          jx0,jy0,jz0,fjx0,fjy0,fjz0,jq0,isaj0;
1065     int              vdwjidx1A,vdwjidx1B;
1066     __m128d          jx1,jy1,jz1,fjx1,fjy1,fjz1,jq1,isaj1;
1067     int              vdwjidx2A,vdwjidx2B;
1068     __m128d          jx2,jy2,jz2,fjx2,fjy2,fjz2,jq2,isaj2;
1069     __m128d          dx00,dy00,dz00,rsq00,rinv00,rinvsq00,r00,qq00,c6_00,c12_00;
1070     __m128d          dx01,dy01,dz01,rsq01,rinv01,rinvsq01,r01,qq01,c6_01,c12_01;
1071     __m128d          dx02,dy02,dz02,rsq02,rinv02,rinvsq02,r02,qq02,c6_02,c12_02;
1072     __m128d          dx10,dy10,dz10,rsq10,rinv10,rinvsq10,r10,qq10,c6_10,c12_10;
1073     __m128d          dx11,dy11,dz11,rsq11,rinv11,rinvsq11,r11,qq11,c6_11,c12_11;
1074     __m128d          dx12,dy12,dz12,rsq12,rinv12,rinvsq12,r12,qq12,c6_12,c12_12;
1075     __m128d          dx20,dy20,dz20,rsq20,rinv20,rinvsq20,r20,qq20,c6_20,c12_20;
1076     __m128d          dx21,dy21,dz21,rsq21,rinv21,rinvsq21,r21,qq21,c6_21,c12_21;
1077     __m128d          dx22,dy22,dz22,rsq22,rinv22,rinvsq22,r22,qq22,c6_22,c12_22;
1078     __m128d          velec,felec,velecsum,facel,crf,krf,krf2;
1079     real             *charge;
1080     int              nvdwtype;
1081     __m128d          rinvsix,rvdw,vvdw,vvdw6,vvdw12,fvdw,fvdw6,fvdw12,vvdwsum,sh_vdw_invrcut6;
1082     int              *vdwtype;
1083     real             *vdwparam;
1084     __m128d          one_sixth   = _mm_set1_pd(1.0/6.0);
1085     __m128d          one_twelfth = _mm_set1_pd(1.0/12.0);
1086     __m128d          dummy_mask,cutoff_mask;
1087     __m128d          signbit   = gmx_mm_castsi128_pd( _mm_set_epi32(0x80000000,0x00000000,0x80000000,0x00000000) );
1088     __m128d          one     = _mm_set1_pd(1.0);
1089     __m128d          two     = _mm_set1_pd(2.0);
1090     x                = xx[0];
1091     f                = ff[0];
1092
1093     nri              = nlist->nri;
1094     iinr             = nlist->iinr;
1095     jindex           = nlist->jindex;
1096     jjnr             = nlist->jjnr;
1097     shiftidx         = nlist->shift;
1098     gid              = nlist->gid;
1099     shiftvec         = fr->shift_vec[0];
1100     fshift           = fr->fshift[0];
1101     facel            = _mm_set1_pd(fr->epsfac);
1102     charge           = mdatoms->chargeA;
1103     krf              = _mm_set1_pd(fr->ic->k_rf);
1104     krf2             = _mm_set1_pd(fr->ic->k_rf*2.0);
1105     crf              = _mm_set1_pd(fr->ic->c_rf);
1106     nvdwtype         = fr->ntype;
1107     vdwparam         = fr->nbfp;
1108     vdwtype          = mdatoms->typeA;
1109
1110     /* Setup water-specific parameters */
1111     inr              = nlist->iinr[0];
1112     iq0              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+0]));
1113     iq1              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+1]));
1114     iq2              = _mm_mul_pd(facel,_mm_set1_pd(charge[inr+2]));
1115     vdwioffset0      = 2*nvdwtype*vdwtype[inr+0];
1116
1117     jq0              = _mm_set1_pd(charge[inr+0]);
1118     jq1              = _mm_set1_pd(charge[inr+1]);
1119     jq2              = _mm_set1_pd(charge[inr+2]);
1120     vdwjidx0A        = 2*vdwtype[inr+0];
1121     qq00             = _mm_mul_pd(iq0,jq0);
1122     c6_00            = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A]);
1123     c12_00           = _mm_set1_pd(vdwparam[vdwioffset0+vdwjidx0A+1]);
1124     qq01             = _mm_mul_pd(iq0,jq1);
1125     qq02             = _mm_mul_pd(iq0,jq2);
1126     qq10             = _mm_mul_pd(iq1,jq0);
1127     qq11             = _mm_mul_pd(iq1,jq1);
1128     qq12             = _mm_mul_pd(iq1,jq2);
1129     qq20             = _mm_mul_pd(iq2,jq0);
1130     qq21             = _mm_mul_pd(iq2,jq1);
1131     qq22             = _mm_mul_pd(iq2,jq2);
1132
1133     /* When we use explicit cutoffs the value must be identical for elec and VdW, so use elec as an arbitrary choice */
1134     rcutoff_scalar   = fr->rcoulomb;
1135     rcutoff          = _mm_set1_pd(rcutoff_scalar);
1136     rcutoff2         = _mm_mul_pd(rcutoff,rcutoff);
1137
1138     sh_vdw_invrcut6  = _mm_set1_pd(fr->ic->sh_invrc6);
1139     rvdw             = _mm_set1_pd(fr->rvdw);
1140
1141     /* Avoid stupid compiler warnings */
1142     jnrA = jnrB = 0;
1143     j_coord_offsetA = 0;
1144     j_coord_offsetB = 0;
1145
1146     outeriter        = 0;
1147     inneriter        = 0;
1148
1149     /* Start outer loop over neighborlists */
1150     for(iidx=0; iidx<nri; iidx++)
1151     {
1152         /* Load shift vector for this list */
1153         i_shift_offset   = DIM*shiftidx[iidx];
1154
1155         /* Load limits for loop over neighbors */
1156         j_index_start    = jindex[iidx];
1157         j_index_end      = jindex[iidx+1];
1158
1159         /* Get outer coordinate index */
1160         inr              = iinr[iidx];
1161         i_coord_offset   = DIM*inr;
1162
1163         /* Load i particle coords and add shift vector */
1164         gmx_mm_load_shift_and_3rvec_broadcast_pd(shiftvec+i_shift_offset,x+i_coord_offset,
1165                                                  &ix0,&iy0,&iz0,&ix1,&iy1,&iz1,&ix2,&iy2,&iz2);
1166
1167         fix0             = _mm_setzero_pd();
1168         fiy0             = _mm_setzero_pd();
1169         fiz0             = _mm_setzero_pd();
1170         fix1             = _mm_setzero_pd();
1171         fiy1             = _mm_setzero_pd();
1172         fiz1             = _mm_setzero_pd();
1173         fix2             = _mm_setzero_pd();
1174         fiy2             = _mm_setzero_pd();
1175         fiz2             = _mm_setzero_pd();
1176
1177         /* Start inner kernel loop */
1178         for(jidx=j_index_start; jidx<j_index_end-1; jidx+=2)
1179         {
1180
1181             /* Get j neighbor index, and coordinate index */
1182             jnrA             = jjnr[jidx];
1183             jnrB             = jjnr[jidx+1];
1184             j_coord_offsetA  = DIM*jnrA;
1185             j_coord_offsetB  = DIM*jnrB;
1186
1187             /* load j atom coordinates */
1188             gmx_mm_load_3rvec_2ptr_swizzle_pd(x+j_coord_offsetA,x+j_coord_offsetB,
1189                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1190
1191             /* Calculate displacement vector */
1192             dx00             = _mm_sub_pd(ix0,jx0);
1193             dy00             = _mm_sub_pd(iy0,jy0);
1194             dz00             = _mm_sub_pd(iz0,jz0);
1195             dx01             = _mm_sub_pd(ix0,jx1);
1196             dy01             = _mm_sub_pd(iy0,jy1);
1197             dz01             = _mm_sub_pd(iz0,jz1);
1198             dx02             = _mm_sub_pd(ix0,jx2);
1199             dy02             = _mm_sub_pd(iy0,jy2);
1200             dz02             = _mm_sub_pd(iz0,jz2);
1201             dx10             = _mm_sub_pd(ix1,jx0);
1202             dy10             = _mm_sub_pd(iy1,jy0);
1203             dz10             = _mm_sub_pd(iz1,jz0);
1204             dx11             = _mm_sub_pd(ix1,jx1);
1205             dy11             = _mm_sub_pd(iy1,jy1);
1206             dz11             = _mm_sub_pd(iz1,jz1);
1207             dx12             = _mm_sub_pd(ix1,jx2);
1208             dy12             = _mm_sub_pd(iy1,jy2);
1209             dz12             = _mm_sub_pd(iz1,jz2);
1210             dx20             = _mm_sub_pd(ix2,jx0);
1211             dy20             = _mm_sub_pd(iy2,jy0);
1212             dz20             = _mm_sub_pd(iz2,jz0);
1213             dx21             = _mm_sub_pd(ix2,jx1);
1214             dy21             = _mm_sub_pd(iy2,jy1);
1215             dz21             = _mm_sub_pd(iz2,jz1);
1216             dx22             = _mm_sub_pd(ix2,jx2);
1217             dy22             = _mm_sub_pd(iy2,jy2);
1218             dz22             = _mm_sub_pd(iz2,jz2);
1219
1220             /* Calculate squared distance and things based on it */
1221             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1222             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1223             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1224             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1225             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1226             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1227             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1228             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1229             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1230
1231             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1232             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1233             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1234             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1235             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1236             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1237             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1238             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1239             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1240
1241             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1242             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1243             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1244             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1245             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1246             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1247             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1248             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1249             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1250
1251             fjx0             = _mm_setzero_pd();
1252             fjy0             = _mm_setzero_pd();
1253             fjz0             = _mm_setzero_pd();
1254             fjx1             = _mm_setzero_pd();
1255             fjy1             = _mm_setzero_pd();
1256             fjz1             = _mm_setzero_pd();
1257             fjx2             = _mm_setzero_pd();
1258             fjy2             = _mm_setzero_pd();
1259             fjz2             = _mm_setzero_pd();
1260
1261             /**************************
1262              * CALCULATE INTERACTIONS *
1263              **************************/
1264
1265             if (gmx_mm_any_lt(rsq00,rcutoff2))
1266             {
1267
1268             /* REACTION-FIELD ELECTROSTATICS */
1269             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
1270
1271             /* LENNARD-JONES DISPERSION/REPULSION */
1272
1273             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1274             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1275
1276             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1277
1278             fscal            = _mm_add_pd(felec,fvdw);
1279
1280             fscal            = _mm_and_pd(fscal,cutoff_mask);
1281
1282             /* Update vectorial force */
1283             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1284             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1285             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1286
1287             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1288             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1289             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1290
1291             }
1292
1293             /**************************
1294              * CALCULATE INTERACTIONS *
1295              **************************/
1296
1297             if (gmx_mm_any_lt(rsq01,rcutoff2))
1298             {
1299
1300             /* REACTION-FIELD ELECTROSTATICS */
1301             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
1302
1303             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
1304
1305             fscal            = felec;
1306
1307             fscal            = _mm_and_pd(fscal,cutoff_mask);
1308
1309             /* Update vectorial force */
1310             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1311             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1312             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1313
1314             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1315             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1316             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1317
1318             }
1319
1320             /**************************
1321              * CALCULATE INTERACTIONS *
1322              **************************/
1323
1324             if (gmx_mm_any_lt(rsq02,rcutoff2))
1325             {
1326
1327             /* REACTION-FIELD ELECTROSTATICS */
1328             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
1329
1330             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
1331
1332             fscal            = felec;
1333
1334             fscal            = _mm_and_pd(fscal,cutoff_mask);
1335
1336             /* Update vectorial force */
1337             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1338             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1339             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1340
1341             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1342             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1343             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1344
1345             }
1346
1347             /**************************
1348              * CALCULATE INTERACTIONS *
1349              **************************/
1350
1351             if (gmx_mm_any_lt(rsq10,rcutoff2))
1352             {
1353
1354             /* REACTION-FIELD ELECTROSTATICS */
1355             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
1356
1357             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
1358
1359             fscal            = felec;
1360
1361             fscal            = _mm_and_pd(fscal,cutoff_mask);
1362
1363             /* Update vectorial force */
1364             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1365             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1366             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1367
1368             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1369             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1370             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1371
1372             }
1373
1374             /**************************
1375              * CALCULATE INTERACTIONS *
1376              **************************/
1377
1378             if (gmx_mm_any_lt(rsq11,rcutoff2))
1379             {
1380
1381             /* REACTION-FIELD ELECTROSTATICS */
1382             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1383
1384             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1385
1386             fscal            = felec;
1387
1388             fscal            = _mm_and_pd(fscal,cutoff_mask);
1389
1390             /* Update vectorial force */
1391             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1392             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1393             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1394
1395             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1396             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1397             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1398
1399             }
1400
1401             /**************************
1402              * CALCULATE INTERACTIONS *
1403              **************************/
1404
1405             if (gmx_mm_any_lt(rsq12,rcutoff2))
1406             {
1407
1408             /* REACTION-FIELD ELECTROSTATICS */
1409             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1410
1411             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1412
1413             fscal            = felec;
1414
1415             fscal            = _mm_and_pd(fscal,cutoff_mask);
1416
1417             /* Update vectorial force */
1418             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1419             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1420             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1421
1422             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1423             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1424             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1425
1426             }
1427
1428             /**************************
1429              * CALCULATE INTERACTIONS *
1430              **************************/
1431
1432             if (gmx_mm_any_lt(rsq20,rcutoff2))
1433             {
1434
1435             /* REACTION-FIELD ELECTROSTATICS */
1436             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
1437
1438             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
1439
1440             fscal            = felec;
1441
1442             fscal            = _mm_and_pd(fscal,cutoff_mask);
1443
1444             /* Update vectorial force */
1445             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1446             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1447             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1448
1449             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1450             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1451             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1452
1453             }
1454
1455             /**************************
1456              * CALCULATE INTERACTIONS *
1457              **************************/
1458
1459             if (gmx_mm_any_lt(rsq21,rcutoff2))
1460             {
1461
1462             /* REACTION-FIELD ELECTROSTATICS */
1463             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1464
1465             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1466
1467             fscal            = felec;
1468
1469             fscal            = _mm_and_pd(fscal,cutoff_mask);
1470
1471             /* Update vectorial force */
1472             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1473             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1474             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1475
1476             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1477             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1478             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1479
1480             }
1481
1482             /**************************
1483              * CALCULATE INTERACTIONS *
1484              **************************/
1485
1486             if (gmx_mm_any_lt(rsq22,rcutoff2))
1487             {
1488
1489             /* REACTION-FIELD ELECTROSTATICS */
1490             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1491
1492             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1493
1494             fscal            = felec;
1495
1496             fscal            = _mm_and_pd(fscal,cutoff_mask);
1497
1498             /* Update vectorial force */
1499             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1500             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1501             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1502
1503             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1504             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1505             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1506
1507             }
1508
1509             gmx_mm_decrement_3rvec_2ptr_swizzle_pd(f+j_coord_offsetA,f+j_coord_offsetB,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1510
1511             /* Inner loop uses 304 flops */
1512         }
1513
1514         if(jidx<j_index_end)
1515         {
1516
1517             jnrA             = jjnr[jidx];
1518             j_coord_offsetA  = DIM*jnrA;
1519
1520             /* load j atom coordinates */
1521             gmx_mm_load_3rvec_1ptr_swizzle_pd(x+j_coord_offsetA,
1522                                               &jx0,&jy0,&jz0,&jx1,&jy1,&jz1,&jx2,&jy2,&jz2);
1523
1524             /* Calculate displacement vector */
1525             dx00             = _mm_sub_pd(ix0,jx0);
1526             dy00             = _mm_sub_pd(iy0,jy0);
1527             dz00             = _mm_sub_pd(iz0,jz0);
1528             dx01             = _mm_sub_pd(ix0,jx1);
1529             dy01             = _mm_sub_pd(iy0,jy1);
1530             dz01             = _mm_sub_pd(iz0,jz1);
1531             dx02             = _mm_sub_pd(ix0,jx2);
1532             dy02             = _mm_sub_pd(iy0,jy2);
1533             dz02             = _mm_sub_pd(iz0,jz2);
1534             dx10             = _mm_sub_pd(ix1,jx0);
1535             dy10             = _mm_sub_pd(iy1,jy0);
1536             dz10             = _mm_sub_pd(iz1,jz0);
1537             dx11             = _mm_sub_pd(ix1,jx1);
1538             dy11             = _mm_sub_pd(iy1,jy1);
1539             dz11             = _mm_sub_pd(iz1,jz1);
1540             dx12             = _mm_sub_pd(ix1,jx2);
1541             dy12             = _mm_sub_pd(iy1,jy2);
1542             dz12             = _mm_sub_pd(iz1,jz2);
1543             dx20             = _mm_sub_pd(ix2,jx0);
1544             dy20             = _mm_sub_pd(iy2,jy0);
1545             dz20             = _mm_sub_pd(iz2,jz0);
1546             dx21             = _mm_sub_pd(ix2,jx1);
1547             dy21             = _mm_sub_pd(iy2,jy1);
1548             dz21             = _mm_sub_pd(iz2,jz1);
1549             dx22             = _mm_sub_pd(ix2,jx2);
1550             dy22             = _mm_sub_pd(iy2,jy2);
1551             dz22             = _mm_sub_pd(iz2,jz2);
1552
1553             /* Calculate squared distance and things based on it */
1554             rsq00            = gmx_mm_calc_rsq_pd(dx00,dy00,dz00);
1555             rsq01            = gmx_mm_calc_rsq_pd(dx01,dy01,dz01);
1556             rsq02            = gmx_mm_calc_rsq_pd(dx02,dy02,dz02);
1557             rsq10            = gmx_mm_calc_rsq_pd(dx10,dy10,dz10);
1558             rsq11            = gmx_mm_calc_rsq_pd(dx11,dy11,dz11);
1559             rsq12            = gmx_mm_calc_rsq_pd(dx12,dy12,dz12);
1560             rsq20            = gmx_mm_calc_rsq_pd(dx20,dy20,dz20);
1561             rsq21            = gmx_mm_calc_rsq_pd(dx21,dy21,dz21);
1562             rsq22            = gmx_mm_calc_rsq_pd(dx22,dy22,dz22);
1563
1564             rinv00           = gmx_mm_invsqrt_pd(rsq00);
1565             rinv01           = gmx_mm_invsqrt_pd(rsq01);
1566             rinv02           = gmx_mm_invsqrt_pd(rsq02);
1567             rinv10           = gmx_mm_invsqrt_pd(rsq10);
1568             rinv11           = gmx_mm_invsqrt_pd(rsq11);
1569             rinv12           = gmx_mm_invsqrt_pd(rsq12);
1570             rinv20           = gmx_mm_invsqrt_pd(rsq20);
1571             rinv21           = gmx_mm_invsqrt_pd(rsq21);
1572             rinv22           = gmx_mm_invsqrt_pd(rsq22);
1573
1574             rinvsq00         = _mm_mul_pd(rinv00,rinv00);
1575             rinvsq01         = _mm_mul_pd(rinv01,rinv01);
1576             rinvsq02         = _mm_mul_pd(rinv02,rinv02);
1577             rinvsq10         = _mm_mul_pd(rinv10,rinv10);
1578             rinvsq11         = _mm_mul_pd(rinv11,rinv11);
1579             rinvsq12         = _mm_mul_pd(rinv12,rinv12);
1580             rinvsq20         = _mm_mul_pd(rinv20,rinv20);
1581             rinvsq21         = _mm_mul_pd(rinv21,rinv21);
1582             rinvsq22         = _mm_mul_pd(rinv22,rinv22);
1583
1584             fjx0             = _mm_setzero_pd();
1585             fjy0             = _mm_setzero_pd();
1586             fjz0             = _mm_setzero_pd();
1587             fjx1             = _mm_setzero_pd();
1588             fjy1             = _mm_setzero_pd();
1589             fjz1             = _mm_setzero_pd();
1590             fjx2             = _mm_setzero_pd();
1591             fjy2             = _mm_setzero_pd();
1592             fjz2             = _mm_setzero_pd();
1593
1594             /**************************
1595              * CALCULATE INTERACTIONS *
1596              **************************/
1597
1598             if (gmx_mm_any_lt(rsq00,rcutoff2))
1599             {
1600
1601             /* REACTION-FIELD ELECTROSTATICS */
1602             felec            = _mm_mul_pd(qq00,_mm_msub_pd(rinv00,rinvsq00,krf2));
1603
1604             /* LENNARD-JONES DISPERSION/REPULSION */
1605
1606             rinvsix          = _mm_mul_pd(_mm_mul_pd(rinvsq00,rinvsq00),rinvsq00);
1607             fvdw             = _mm_mul_pd(_mm_msub_pd(c12_00,rinvsix,c6_00),_mm_mul_pd(rinvsix,rinvsq00));
1608
1609             cutoff_mask      = _mm_cmplt_pd(rsq00,rcutoff2);
1610
1611             fscal            = _mm_add_pd(felec,fvdw);
1612
1613             fscal            = _mm_and_pd(fscal,cutoff_mask);
1614
1615             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1616
1617             /* Update vectorial force */
1618             fix0             = _mm_macc_pd(dx00,fscal,fix0);
1619             fiy0             = _mm_macc_pd(dy00,fscal,fiy0);
1620             fiz0             = _mm_macc_pd(dz00,fscal,fiz0);
1621
1622             fjx0             = _mm_macc_pd(dx00,fscal,fjx0);
1623             fjy0             = _mm_macc_pd(dy00,fscal,fjy0);
1624             fjz0             = _mm_macc_pd(dz00,fscal,fjz0);
1625
1626             }
1627
1628             /**************************
1629              * CALCULATE INTERACTIONS *
1630              **************************/
1631
1632             if (gmx_mm_any_lt(rsq01,rcutoff2))
1633             {
1634
1635             /* REACTION-FIELD ELECTROSTATICS */
1636             felec            = _mm_mul_pd(qq01,_mm_msub_pd(rinv01,rinvsq01,krf2));
1637
1638             cutoff_mask      = _mm_cmplt_pd(rsq01,rcutoff2);
1639
1640             fscal            = felec;
1641
1642             fscal            = _mm_and_pd(fscal,cutoff_mask);
1643
1644             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1645
1646             /* Update vectorial force */
1647             fix0             = _mm_macc_pd(dx01,fscal,fix0);
1648             fiy0             = _mm_macc_pd(dy01,fscal,fiy0);
1649             fiz0             = _mm_macc_pd(dz01,fscal,fiz0);
1650
1651             fjx1             = _mm_macc_pd(dx01,fscal,fjx1);
1652             fjy1             = _mm_macc_pd(dy01,fscal,fjy1);
1653             fjz1             = _mm_macc_pd(dz01,fscal,fjz1);
1654
1655             }
1656
1657             /**************************
1658              * CALCULATE INTERACTIONS *
1659              **************************/
1660
1661             if (gmx_mm_any_lt(rsq02,rcutoff2))
1662             {
1663
1664             /* REACTION-FIELD ELECTROSTATICS */
1665             felec            = _mm_mul_pd(qq02,_mm_msub_pd(rinv02,rinvsq02,krf2));
1666
1667             cutoff_mask      = _mm_cmplt_pd(rsq02,rcutoff2);
1668
1669             fscal            = felec;
1670
1671             fscal            = _mm_and_pd(fscal,cutoff_mask);
1672
1673             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1674
1675             /* Update vectorial force */
1676             fix0             = _mm_macc_pd(dx02,fscal,fix0);
1677             fiy0             = _mm_macc_pd(dy02,fscal,fiy0);
1678             fiz0             = _mm_macc_pd(dz02,fscal,fiz0);
1679
1680             fjx2             = _mm_macc_pd(dx02,fscal,fjx2);
1681             fjy2             = _mm_macc_pd(dy02,fscal,fjy2);
1682             fjz2             = _mm_macc_pd(dz02,fscal,fjz2);
1683
1684             }
1685
1686             /**************************
1687              * CALCULATE INTERACTIONS *
1688              **************************/
1689
1690             if (gmx_mm_any_lt(rsq10,rcutoff2))
1691             {
1692
1693             /* REACTION-FIELD ELECTROSTATICS */
1694             felec            = _mm_mul_pd(qq10,_mm_msub_pd(rinv10,rinvsq10,krf2));
1695
1696             cutoff_mask      = _mm_cmplt_pd(rsq10,rcutoff2);
1697
1698             fscal            = felec;
1699
1700             fscal            = _mm_and_pd(fscal,cutoff_mask);
1701
1702             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1703
1704             /* Update vectorial force */
1705             fix1             = _mm_macc_pd(dx10,fscal,fix1);
1706             fiy1             = _mm_macc_pd(dy10,fscal,fiy1);
1707             fiz1             = _mm_macc_pd(dz10,fscal,fiz1);
1708
1709             fjx0             = _mm_macc_pd(dx10,fscal,fjx0);
1710             fjy0             = _mm_macc_pd(dy10,fscal,fjy0);
1711             fjz0             = _mm_macc_pd(dz10,fscal,fjz0);
1712
1713             }
1714
1715             /**************************
1716              * CALCULATE INTERACTIONS *
1717              **************************/
1718
1719             if (gmx_mm_any_lt(rsq11,rcutoff2))
1720             {
1721
1722             /* REACTION-FIELD ELECTROSTATICS */
1723             felec            = _mm_mul_pd(qq11,_mm_msub_pd(rinv11,rinvsq11,krf2));
1724
1725             cutoff_mask      = _mm_cmplt_pd(rsq11,rcutoff2);
1726
1727             fscal            = felec;
1728
1729             fscal            = _mm_and_pd(fscal,cutoff_mask);
1730
1731             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1732
1733             /* Update vectorial force */
1734             fix1             = _mm_macc_pd(dx11,fscal,fix1);
1735             fiy1             = _mm_macc_pd(dy11,fscal,fiy1);
1736             fiz1             = _mm_macc_pd(dz11,fscal,fiz1);
1737
1738             fjx1             = _mm_macc_pd(dx11,fscal,fjx1);
1739             fjy1             = _mm_macc_pd(dy11,fscal,fjy1);
1740             fjz1             = _mm_macc_pd(dz11,fscal,fjz1);
1741
1742             }
1743
1744             /**************************
1745              * CALCULATE INTERACTIONS *
1746              **************************/
1747
1748             if (gmx_mm_any_lt(rsq12,rcutoff2))
1749             {
1750
1751             /* REACTION-FIELD ELECTROSTATICS */
1752             felec            = _mm_mul_pd(qq12,_mm_msub_pd(rinv12,rinvsq12,krf2));
1753
1754             cutoff_mask      = _mm_cmplt_pd(rsq12,rcutoff2);
1755
1756             fscal            = felec;
1757
1758             fscal            = _mm_and_pd(fscal,cutoff_mask);
1759
1760             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1761
1762             /* Update vectorial force */
1763             fix1             = _mm_macc_pd(dx12,fscal,fix1);
1764             fiy1             = _mm_macc_pd(dy12,fscal,fiy1);
1765             fiz1             = _mm_macc_pd(dz12,fscal,fiz1);
1766
1767             fjx2             = _mm_macc_pd(dx12,fscal,fjx2);
1768             fjy2             = _mm_macc_pd(dy12,fscal,fjy2);
1769             fjz2             = _mm_macc_pd(dz12,fscal,fjz2);
1770
1771             }
1772
1773             /**************************
1774              * CALCULATE INTERACTIONS *
1775              **************************/
1776
1777             if (gmx_mm_any_lt(rsq20,rcutoff2))
1778             {
1779
1780             /* REACTION-FIELD ELECTROSTATICS */
1781             felec            = _mm_mul_pd(qq20,_mm_msub_pd(rinv20,rinvsq20,krf2));
1782
1783             cutoff_mask      = _mm_cmplt_pd(rsq20,rcutoff2);
1784
1785             fscal            = felec;
1786
1787             fscal            = _mm_and_pd(fscal,cutoff_mask);
1788
1789             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1790
1791             /* Update vectorial force */
1792             fix2             = _mm_macc_pd(dx20,fscal,fix2);
1793             fiy2             = _mm_macc_pd(dy20,fscal,fiy2);
1794             fiz2             = _mm_macc_pd(dz20,fscal,fiz2);
1795
1796             fjx0             = _mm_macc_pd(dx20,fscal,fjx0);
1797             fjy0             = _mm_macc_pd(dy20,fscal,fjy0);
1798             fjz0             = _mm_macc_pd(dz20,fscal,fjz0);
1799
1800             }
1801
1802             /**************************
1803              * CALCULATE INTERACTIONS *
1804              **************************/
1805
1806             if (gmx_mm_any_lt(rsq21,rcutoff2))
1807             {
1808
1809             /* REACTION-FIELD ELECTROSTATICS */
1810             felec            = _mm_mul_pd(qq21,_mm_msub_pd(rinv21,rinvsq21,krf2));
1811
1812             cutoff_mask      = _mm_cmplt_pd(rsq21,rcutoff2);
1813
1814             fscal            = felec;
1815
1816             fscal            = _mm_and_pd(fscal,cutoff_mask);
1817
1818             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1819
1820             /* Update vectorial force */
1821             fix2             = _mm_macc_pd(dx21,fscal,fix2);
1822             fiy2             = _mm_macc_pd(dy21,fscal,fiy2);
1823             fiz2             = _mm_macc_pd(dz21,fscal,fiz2);
1824
1825             fjx1             = _mm_macc_pd(dx21,fscal,fjx1);
1826             fjy1             = _mm_macc_pd(dy21,fscal,fjy1);
1827             fjz1             = _mm_macc_pd(dz21,fscal,fjz1);
1828
1829             }
1830
1831             /**************************
1832              * CALCULATE INTERACTIONS *
1833              **************************/
1834
1835             if (gmx_mm_any_lt(rsq22,rcutoff2))
1836             {
1837
1838             /* REACTION-FIELD ELECTROSTATICS */
1839             felec            = _mm_mul_pd(qq22,_mm_msub_pd(rinv22,rinvsq22,krf2));
1840
1841             cutoff_mask      = _mm_cmplt_pd(rsq22,rcutoff2);
1842
1843             fscal            = felec;
1844
1845             fscal            = _mm_and_pd(fscal,cutoff_mask);
1846
1847             fscal            = _mm_unpacklo_pd(fscal,_mm_setzero_pd());
1848
1849             /* Update vectorial force */
1850             fix2             = _mm_macc_pd(dx22,fscal,fix2);
1851             fiy2             = _mm_macc_pd(dy22,fscal,fiy2);
1852             fiz2             = _mm_macc_pd(dz22,fscal,fiz2);
1853
1854             fjx2             = _mm_macc_pd(dx22,fscal,fjx2);
1855             fjy2             = _mm_macc_pd(dy22,fscal,fjy2);
1856             fjz2             = _mm_macc_pd(dz22,fscal,fjz2);
1857
1858             }
1859
1860             gmx_mm_decrement_3rvec_1ptr_swizzle_pd(f+j_coord_offsetA,fjx0,fjy0,fjz0,fjx1,fjy1,fjz1,fjx2,fjy2,fjz2);
1861
1862             /* Inner loop uses 304 flops */
1863         }
1864
1865         /* End of innermost loop */
1866
1867         gmx_mm_update_iforce_3atom_swizzle_pd(fix0,fiy0,fiz0,fix1,fiy1,fiz1,fix2,fiy2,fiz2,
1868                                               f+i_coord_offset,fshift+i_shift_offset);
1869
1870         /* Increment number of inner iterations */
1871         inneriter                  += j_index_end - j_index_start;
1872
1873         /* Outer loop uses 18 flops */
1874     }
1875
1876     /* Increment number of outer iterations */
1877     outeriter        += nri;
1878
1879     /* Update outer/inner flops */
1880
1881     inc_nrnb(nrnb,eNR_NBKERNEL_ELEC_VDW_W3W3_F,outeriter*18 + inneriter*304);
1882 }